Add Blackwell MXAttention scaled-MMA backend

[yiliu30]

Summary

• Add the MXAttention UOS + PNQ MXFP4 backend for vLLM-Omni.
• Use native Triton tl.dot_scaled for MXFP4 QK and split-PV MMA.
• Add the SM100 Python workaround for Triton's accumulator-init/TMEM compiler issue.
• Add reference quantization utilities, tests, README documentation, and an implementation handoff.

Root cause and fix

On Blackwell, Triton's accumulator-init optimization can rewrite a loop-carried scaled-MMA accumulator into an invalid immutable TMEM allocation. The runtime workaround skips that pass only for SM100 compilation. The optimized PV path uses four 128x32 accumulators because a single 128x128 scaled-MMA layout is rejected by Blackwell legalization.

Validation

• Triton 3.7.1, Torch 2.11.0+cu130.
• GPU3 / NVIDIA B200 native QK and PV tl.dot_scaled probes.
• Causal, noncausal, padded, and multi-block attention cases.
• Public MXAttention API with allow_fallback=False.
• Wan2.2 TI2V-5B E2E with explicit SAGE_ATTN: model-card configuration 1280x704, 121 frames, 50 steps, 24 FPS; completed successfully in approximately 260.5 seconds.
• MP4 output validated at 121 frames, 1280x704, 24 FPS.

See docs/MXATTENTION_HANDOFF.md for runtime configuration and follow-up work.