feat(projection_kernel): add attention-head subspace affinity

[janmenjayap]

Description

Implements #1720.

Adds a basis-invariant Projection Kernel analysis surface in two reviewable commits:

1. model-independent reduced-SVD basis extraction, principal angles, raw and normalized PK,
   numerical-rank metadata, roundoff-bound handling, and random-subspace moments;
2. a TransformerBridge OQ/OK/OV head-affinity wrapper with native GQA KV-head identity,
   hybrid-layer indices, forward/all layer masks, bounded-memory tiled scoring, and ranking.

The implementation promotes fp16/bf16 inputs to fp32 for stable SVDs, preserves float64,
detaches model weights from autograd, and reports role/layer/head context for malformed or
rank-deficient weights. The accompanying guide documents orientation, rank semantics,
limitations, GQA behavior, and the distinction from Composition Score, with runnable numerical
and GPT-2 TransformerBridge examples.

Validation

• Focused PK suite: 67 passed.
• make check-format: passed.
• uv run mypy .: passed (392 source files).
• uv run build-docs: passed with no PK-specific warnings.
• uv build: source distribution and wheel built successfully.
• Local PR test surfaces (make test-pr did not exit successfully because of the three
  integration failures described below):
  • unit: 5,363 passed;
  • docstring: 18 passed;
  • acceptance: 209 passed;
  • integration: 1,394 passed, including all PK tests; two gated Gemma loads failed because the
    local HF token lacks access, and one existing GraniteMoeHybrid MPS parity test reproduced the
    investigated PyTorch 2.10 MPS numerical divergence.
• GitHub CI: all applicable compatibility, coverage, formatting, typing, docstring, benchmark,
  and notebook checks passed.

Checklist

• I have read and understood the contribution guidelines.
• I have added tests that prove my fix is effective or that my feature works.
• I have added necessary documentation (if appropriate).
• My changes generate no new warnings.
• No changelog file is required by this repository; the user-facing change is documented in
  docs/source/content/projection_kernel.md.

Type of change

• Bug fix
• New feature
• Breaking change
• Documentation update