fix: complete TrainerRank model runtime integration

[bradhilton]

Summary

Completes the TrainerRank integration with the model-runtime behavior added by #758.

• Canonicalize model-specific internal LoRA padding when loading dynamic slots and zero padded slot gradients before optimizer updates, without mutating unrelated slots that may retain backward graphs.
• Pass packed positions, sliding-window variants, model state, and GDN planner configuration through the CP=1 prefix-tree path.
• Configure HybridEP capacity and the exact per-rank model-row extent for CP=1 and CP>1 packed forwards.
• Reject simultaneous TrainerRank data parallelism and expert parallelism before HybridEP collectives. The supported/recommended MoE topology is DP=1 with CP=EP=world_size.

Validation

• uv run --no-sync prek run --all-files
• uv run --no-sync python dev/trainer_rank_fast_check.py -q (75 passed, 4 skipped)
• 2xH200 dependency gate: all 8 Megatron integration tests, including packed attention, depth>1 GDN parity/trainability, TP LoRA gradients, dynamic-slot recompute/stepping, and TP head backward
• 2xH200 CP=2 matrix: all 16 output combinations at depths 0 and 4 with chunk sizes 17 and 8192; mean_abs_pct=0.00155049, head-backward mean_abs_pct=0.0
• 4xH200 unsupported-topology regression: the formerly hanging CP=1/EP=4/DP=4 unequal-tail case raises the documented error on all four ranks before model collectives
• 4xH200 supported-topology regression: Qwen/Qwen3.5-35B-A3B, CP=4/EP=4/DP=1, unequal dynamic slots, retained graphs, backward, and optimizer step
• GPT-OSS CP=1 dynamic-slot forward/backward/optimizer smoke
• Open-ended adversarial reviews from Claude and a separate Codex subagent found no remaining blocker

No files under src/art/megatron/** are changed.