[Kernel][SM70] Add native-g32 AWQ QPN M1 operator

[Leonccaa]

Purpose and stack

Layer 1/3: native operator, binding, independent GPU tests and design contract.
Base is the existing grouped-decode fork draft #8, not a second copy of it.
This layer does not select the operator in any model runtime or enable a flag.

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

• Add a model-specific SM70 Qwen3.8 TP4/native-g32/E512/top-k10 AWQ M=1
  W13/SwiGLU and W2/router-reduction operator using the existing QPN dataflow.
• Consume current prepared banks and caller-owned output/intermediate buffers.
  Support scalar 3-byte and existing 4-byte metadata, without a new weight copy.
• Preserve FP16 scale/bias dequantization, gate/up/SwiGLU boundaries,
  per-route FP16 materialization and original router accumulation order.
  CTA-local FP32 reduction order changes; legacy is not a mathematical oracle.
• Validate shape, dtype, device, contiguity, alignment and non-overlap. Invalid
  expert IDs contribute zero and duplicate IDs keep separate route weights.
• Register the native operator and fake implementation; no sidecar loader.
  The retained LICENSE.v100-skinny identifies the QPN dataflow provenance.

The subsequent layers are default-off runtime admission and independent
cooperative 3-byte W13 metadata loading. Four-byte full-model performance is
attributable to operator plus runtime, not to the third layer.

Scope and duplicate check

Open-PR searches in this fork, 1CatAI/1Cat-vLLM and vllm-project/vllm for
AWQ/QPN/SM70 found no competing native AWQ M1 operator. The QPN search hits
1Cat 1CatAI#349 (PP2/TP4 FP8 tactic stabilization) and 1CatAI#509 (conversion-cache release)
are different work. Merged 1CatAI#477 separates prefill/decode compilation; this
operator adds AWQ expert computation, not a replacement graph framework.
The separate QSA page4 fix used in both full-model validation arms is not
included here. There is no corresponding vllm-project/vllm issue submission.

Test plan and results

Changed-file pre-commit run --files ... and git diff --check: passed.
The three extra parenthesizations of existing NVFP4 fake lambdas in
_sm70_ops.py are required by the pinned formatter, with no semantic change.

On the native SM70 scalar build, the following portable test invocation passed
2/2 (the repository test directory was mounted as /tests):

/evidence/test-venv/bin/python -m pytest -q --confcutdir=/tests \
  /tests/test_sm70_awq_qpn_m1.py

It constructs both prepared layouts independently, checks one-hot boundary
reads and an FP64 W2 reference, changing graph inputs, invalid/duplicate IDs,
alias/alignment/dtype rejection, and the fake op. CPU skips are not GPU passes.

The archived real-checkpoint harness was also run:

/evidence/test-venv/bin/python /native-tools/probe_native.py \
  --manifest /output/native-manifest.json --output /output/probe

• 256 native/prototype/dynamic-graph exact comparisons across both metadata
  layouts and all four logical TP partitions.
• 72 independent FP64 stage checks, no retained rounding-bound violations;
  48 invalid argument rejections; invalid routes zero; fullgraph eager compile.
• One tiny subnormal case remains above the old relative-only smoke threshold
  (relative L2 0.00201759, max absolute 5.96046e-8). It is recorded, not renamed
  a pass under that old rule. The independent reference includes an explicit
  FP16 subnormal rounding floor; it is not an assertion of token identity.
• Scalar native core SHA-256:
  109e8e68852c701a638f05e5574f64f9d040f41497a6696a6320457285f0b434.
• Evidence: qwen38-c1-qpn-awq-20260905/layer-gpu-scalar-r1, exit0/OOMfalse.

These checks establish the tested operator contract, not general full-model
quality. Runtime evidence and observed task-level changes belong to layer 2.

Review status

AI assistance: OpenAI Codex assisted with implementation, analysis, tests and
documentation. Leon directed the scope and requested submission to this fork.
This is a fork review draft; final human line-by-line review and test
ownership are not claimed complete. No upstream submission, default route
change or production deployment is made here.