[Perf][SM70] Coalesce AWQ QPN W13 compact metadata loads

[Leonccaa]

Purpose and stack

Layer 3/3, based on perf/sm70-awq-qpn-runtime-20260905. This isolates the
cooperative 3-byte W13 metadata reader from the native operator (layer1)
and default-off runtime admission (layer2).

Review order: #8 (prerequisite) -> #10 (kernel) -> #12 (runtime) -> #11 (3B loads).

• Read each 96-byte metadata tile with 24 aligned 32-bit loads per warp and
  distribute its 3-byte records with shuffles.
• The final record at byte93 neither reads nor shuffles beyond the tile.
• W2 keeps the scalar 3-byte reader. The 4-byte path is unchanged.
• No metadata representation change, repack, added weight copy or buffer.
  Bias reconstruction, MMA/reduction order, and FP16 boundaries are unchanged.
• Removing this commit restores the scalar 3-byte path. The overall runtime
  remains opt-in; no new flag or automatic layout switch is introduced.

This is not a separate general QPN framework or a duplicate of fork #8,
upstream 1CatAI#477, 1CatAI#349, or 1CatAI#509. Open SM70/AWQ/QPN PR checks found no competing
AWQ compact-metadata reader. Layer1 deliberately retains the scalar reader
so this change can be reviewed and reverted independently.

Test plan and results

The extracted CUDA source is byte-identical to the already GPU-validated
full native source, SHA-256:
55010c76911ba3e6ac6ed4017a6e624dfc3db2d1000221a28df6501d4c9481b5.
The repository portable GPU test also matches the archived test exactly.
Changed-file pre-commit hooks and git diff --check: passed.

Both scalar and cooperative builds were tested with:

/evidence/test-venv/bin/python -m pytest -q --confcutdir=/tests \
  /tests/test_sm70_awq_qpn_m1.py
/evidence/test-venv/bin/python /native-tools/probe_native.py \
  --manifest /output/native-manifest.json --output /output/probe
/evidence/test-venv/bin/python /native-tools/benchmark_layer_split.py \
  --output /output/split-timing.json

Per build: 2 portable GPU tests, 256 real-weight/dynamic-graph exact
comparisons to the same prototype, 72 FP64 stage checks with zero retained
rounding-bound violations, and 48 invalid argument rejections. Both
metadata layouts and all four logical TP partitions were covered. The
tiny-subnormal relative-only smoke exception is retained in the layer1 report;
scalar/cooperative/prototype outputs were exact for that case too.

A single-layer graph probe measured 3-byte scalar21.5280us versus
cooperative20.6160us (about4.2% lower). This is a component observation, not
an end-to-end speedup or statistical throughput claim. Full-model speed and
quality evidence in layer2 uses 4-byte metadata, so it must not be cited
as model-level validation of this 3-byte path.

Evidence: qwen38-c1-qpn-awq-20260905/layer-gpu-{scalar,full}-r1.
Full native core SHA-256:
e5083c1f26392873460c4a798244f454c6adef4acb66cc7d6573d4da8f8916e1.
GPU jobs exited0/OOMfalse; logs/inspect were retained and GPU leases released.
This extraction did not introduce new arithmetic or collect new GPU timings.

AI assistance: OpenAI Codex assisted with implementation, analysis, tests and
documentation. Leon directed scope and requested submission to this fork.
This is a fork review draft, with final human review pending. No upstream
submission, production deployment or broad quality acceptance is claimed.