Sync DeepSeek upstream through V4.1 kernels (#221)

[zyongye]

Summary

• Sync deepseek-ai/FlashMLA:main through commit 07a1089857b63e74e3133630c02b083b75e8d4b2, including the DeepSeek V4.1 kernels from Add kernels for DeepSeek v4.1 deepseek-ai/FlashMLA#221 and the reorganized csrc/kernels tree. The upstream history is represented by one signed sync commit for the target repository's DCO policy.
• Preserve the vLLM fork's stable-ABI registration, SM90 dense-FP8 extension, optional output buffers, thread-safe device properties, and split-workspace optimization.
• Port the V3.2 NVFP4-NoPE/fp8-RoPE sparse-decode path to the upstream layout without regressing V4/V4.1 dispatch.
• Keep repository-local virtual environments out of version control.

Validation

• Full CUDA extension build with CUDA 13.1 for SM90a, SM100a, and SM103a.
• 6 focused pytest checks for stable operator registration, NVFP4 wire format, and optional output forwarding.
• GB200 correctness coverage for:
  • V3.2 FP8 split and no-split decode
  • V4 FP8 decode
  • V4.1 FP8 plus V4.1 FP4 extra-cache decode
  • NVFP4 decode with 64/128 heads, top-k lengths, attention sinks, invalid indices, zero-length requests, and an extra cache
  • real preallocated output buffers for decode and sparse prefill
• Standalone compile/load/registration check for the preserved vLLM SM90 dense-FP8 extension.
• Built and installed flash_mla-1.0.0+29b0d03-cp310-abi3-linux_aarch64.whl into .venv; out-of-tree import registered all 9 operators.

Test dependency note

The upstream fused end-to-end reference script was not run locally because its optional tile_kernels and deep_gemm dependencies are not installed. All fused V4/V4.1 sources were included in the successful full extension build.