Experimental performance family (default-off): load-balance infrastructure, active-box windowing, block-structured AMR, hybrid WENO/Riemann sensors

[sbryngelson]

Summary

This PR adds an opt-in, default-off family of performance features and the measurement
infrastructure they rest on. With all flags at their defaults the only touched production
path is s_mpi_decompose_computational_domain, refactored to compute its equal split
through the new m_box module (byte-identical; covered by the existing suite).

Load-balance infrastructure (common + sim):

• m_box: t_box + partition arithmetic; shared by the decomposer, AMR, and the
  weighted splitter.
• m_load_weight + load_weight_wrt: per-cell load-weight field (active-box, EL-bubble,
  IB, phase-change Newton-iteration contributors) with field output and a per-rank
  imbalance metric.
• m_sfc_partition + sfc_partition_wrt: Morton-SFC tile ordering and chains-on-chains
  balanced partition, reported as a predicted-imbalance diagnostic.
• m_load_balance + load_balance: experimental weighted static Cartesian decomposition
  at init (requires parallel_io), with a min-cells feasibility floor and, when AMR is
  on, fine-work-aware weighting with a deterministic feasibility clamp.
• m_rank_timing + rank_time_wrt: per-rank compute-time diagnostic (halo exchange
  excluded; device-synced on GPU).

Active-box windowing (sim):

• m_active_box + active_box: restricts reconstruction/Riemann/RK windows to a
  light-cone-grown box around non-ambient flow; a debug tripwire guards under-growth.
  Golden-tested (ECABA006) to stay a strict subset while matching the full-domain
  solution.

Block-structured AMR (sim):

• m_amr + m_amr_registers + amr: a two-level 2:1 refined hierarchy with
  conservative restriction / conservative-linear prolongation, per-stage flux registers
  with Berger–Colella refluxing, gradient-based dynamic regrid (amr_regrid_int,
  amr_tag_eps, amr_buf), optional dt/2 subcycling (amr_subcycle), multi-rank
  operation with a mirror-decomposed fine level (patches may span rank boundaries; fine
  halo exchange; distributed flux registers; rank-local regrid), and GPU builds
  (device-resident fine fields and registers, on-device ghost fill/RK/restriction).
  Requires WENO, SSP-RK3, model_eqns=2, single fluid (checker-enforced).

Hybrid reconstruction/flux sensors (sim):

• hybrid_weno (+hybrid_weno_eps): linear-optimal reconstruction in smooth cells, full
  WENO only at flagged discontinuities (Jameson-type density+pressure sensor,
  stencil-dilated, halo-aware).
• hybrid_riemann (+hybrid_smooth_flux): cheap central/Rusanov flux in smooth cells,
  full HLLC at discontinuities (5- and 6-equation blocks).

Motivation

Measured rank imbalance on heterogeneous-cost workloads (bubbles, IB, phase change)
motivates first-class measurement tools; the active box and hybrid sensors give direct
speedups on localized-flow / mostly-smooth cases; AMR concentrates resolution where the
flow needs it, and the load-balance coupling keeps the refined work spread across ranks.

Testing

• Four new golden tests: 5ECBB926 (AMR static patch), 1CBACEB5 (AMR dynamic regrid),
  852CCB81 (AMR subcycling), ECABA006 (active_box 3D strict-subset).
• Locally verified on gfortran/CPU: three-target build; full precheck; the 4 new goldens
  • 4 existing goldens (incl. both periodic-IBM cases, exercising the Fix periodic ib issues #1618 merge) all
    pass; load_balance+amr np=2 end-to-end smoke produces the analytically predicted
    weighted offsets and completes; amr np=2 spanning-patch run completes.
• GPU: nvfortran OpenACC build of the simulation target (the GPU-accelerated
  executable) verified locally (see PR checks for the full matrix).
• All parameters default off; case_validator entries, case.md docs, and
  module_categories are included.

Known-untested configurations

Delegated to CI: Cray ftn, Intel ifx, AMD flang, OpenMP target offload, single/mixed
precision. Hybrid WENO/Riemann ship without a dedicated golden case (flagging for
reviewer judgment; the sensors are default-off and checker-guarded).

Review guide

The 75 commits are arc-ordered and cleanly arc-separable — reviewing by arc is much
easier than by file:

1. 2760da7d…2bb5fdc4 active-box (11)
2. bbf6b2a9…14b837c6 load-weight field + contributors (8)
3. 0161fac0…2795e266 SFC partition diagnostic (6)
4. 6df9c1f0…c43c02a5 weighted decomposition (load_balance) (8)
5. 95398eb3…cc7882d1 rank timing (4)
6. 21c60ffa…5082b535 hybrid WENO/Riemann (10)
7. 74b58771…de244407 m_box refactor + validation hygiene (4)
8. 352f564e…03b59516 AMR: static hierarchy → restriction/prolongation → fine advance →
   refluxing → regrid → subcycling → multi-rank → GPU → mirror decomposition →
   load-balance coupling (20)
9. a1a7e3ad merge of upstream/master (num_procs_x/y/z promotion adopted from Fix periodic ib issues #1618)

Addendum: features added after the initial draft

• Multi-fluid AMR (5-equation multi-component): per-fluid conservative reflux (per-fluid mass defects ~1e-15 through refluxed+subcycled+regridding advance), sum-preserving volume-fraction prolongation (n−1 fractions + closure), mpp_lim required for num_fluids > 1; shock–material-interface demo validated. Known bounded limitation: alpha-sum deviation up to ~5.7e-3 at coarse cells historically hosting a patch face during shock crossing (non-growing, mpp_lim-damped; the volume-fraction K-term is deliberately not refluxed — it is non-conservative).
• 3D validation: free-stream exact (0.0) in 3D with subcycling+regrid armed; 3D blast regrid-tracked at ~1e-14 defects; np=2 seam-spanning element-exact; code audit found no y/z asymmetry.
• Golden coverage is now 5 cases: 1D static / 1D regrid / 1D subcycle / 3D static / 1D two-fluid.

Further additions

• AMR restart (SP10): fine-level save/restore with regridded-box persistence, both IO modes (serial per-rank + parallel MPI-IO); restart continuation element-exact (save-then-restart == straight run), same-num_procs required (np-flexible restart is future work).
• Viscous AMR (SP11): viscous prohibition lifted; viscous stress/work refluxed through the existing registers (enters rhs as a flux_src_n face-flux difference, same form as advective flux) so coarse/fine boundaries match total flux; energy conservation 0.0, accuracy triplet coarse 2.49e-4 ≫ two-level 6.89e-5 ≈ fine 5.04e-5. A fine-ghost-coordinate bug (viscous gradient using stale coarse dx at the fine subdomain/patch edge — invisible to WENO, which uses only interior dx) was found by an np=2 exactness probe and fixed; the fine viscous seam is now byte-exact across ranks. Residual: a bounded (~1e-6) np-dependence remains only at the coarse/fine patch boundary from prolongation-derived ghost gradients (AMR's inherently-approximate coupling zone); the density-gradient tagger senses shear poorly (buffered/static patch recommended; error-estimator taggers are future work).

Multi-block AMR + terminology

• Multi-block AMR (SP12a): tagged cells are clustered (Berger–Rigoutsos + a min-separation merge) into a LIST of separated refinement blocks, so multiple separated features (e.g. a shock and a contact that have separated) each get their own tight block instead of one bounding box wastefully spanning both plus the smooth gap between — measured 66% fewer cells refined on a two-interface Sod, conservation and np=1==np=2 element-exactness preserved. amr_max_blocks (default 4; N fixed-size slots, ~N× device memory — compute efficiency is the goal, memory efficiency a follow-up), amr_cluster_eff (default 0.7). Fine blocks stay ≥ buff_size apart ⇒ no fine–fine coupling; all existing per-block machinery (multi-rank, GPU, subcycle, viscous, multi-fluid) loops over the block list unchanged.
• Terminology: AMR's refinement regions are now called blocks (amr_block_beg/end, amr_max_blocks) — disambiguated from MFC's initial-condition patch_icpp. (Draft-stage rename; golden values unchanged.)