feat(sm100): add FP4 activation support

[mgoin]

Adds native SM100-family E2M1 FP4 activations for MXFP4 (group32/E8M0) and NVFP4 (group16/E4M3), with BF16 outputs. Dense, indexed, grouped-contiguous and grouped-masked GEMMs reuse existing weight packing and output handling. Coverage includes supported smaller integer weights dequantized to FP4, channel scales and NVFP4 secondary tensor scales.

Packed FP4 activations stay in shared memory; canonical weight fragments and scales are staged into TMEM for native block-scaled UMMA. The initial schedule uses BM64/BM128 and K128/K256 stages with bounded occupancy choices. Existing integer A4 and SM120 execution retain their paths. Explicit input scales and indexed gathers use the physical packed row stride, including logical K tails.

Integration status: this is the independently validated A4 branch against main. It overlaps A8 #73 in the new MXUMMA executor, eligibility and geometry selection. Keep it draft until those shared changes are reconciled; the two drafts cannot simply be merged as written. The A8 native-output optimization #74 has not been ported here. This preserves the tested A4 implementation for review without claiming a validated combined A4/A8 executor.

11 files, +564/-15, including 295 added lines in the existing MXMMA test module. The publication tree is identical to validated 83c616a. Open Humming PRs were checked; current SM90 tuning work does not duplicate FP4 activation support on SM100.

Validation on B300/SM103, PyTorch 2.13.0+cu130, CUDA 13.0, generated sm_100f (CUDA >=12.9):

CUDA_VISIBLE_DEVICES=3 PYTHONPATH="$PWD" /home/mgoin/code/vllm/.venv/bin/python -m pytest tests/kernels/humming/test_mxmma.py -k sm100 -q
CUDA_VISIBLE_DEVICES=3 PYTHONPATH="$PWD" /home/mgoin/code/vllm/.venv/bin/python -m pytest tests/kernels/humming/test_umma.py -k pipeline_stage_reuse -q
CUDA_VISIBLE_DEVICES=2 PYTHONPATH="$PWD" /home/mgoin/code/vllm/.venv/bin/python -m pytest tests/kernels/humming/test_mxmma.py -k 'public_layer_reuses_packed_weights and 2880' -q -x

The support campaign passed 20 native cases across its recorded runs, public internal/prequantized input tests, compiler eligibility checks, five BF16 pipeline regressions and eight legacy integer-A4 comparisons. The later scale-stride correction passed 24 strict cases, six tensor-operation memchecks with zero errors and both retained K2880 public regressions. All active outputs are explicitly checked against independent references. Ruff and git diff --check passed.

Identical-input dense MXFP4 projection, N=K=4096, CUPTI/CUDA graphs/cold L2: at M512, Humming 1201 TFLOP/s versus DeepGEMM 1588; at M2048, 1786 versus 2818. Quantization, packing and routing are excluded. This is initial kernel/library support, not native performance parity; no A4 vLLM model evaluation or distributed EP measurement was performed.

AI assistance: implementation, review and validation used OpenAI Codex, including GPT-6 Astra.