[Feature] Add optional FlashMLA backend for DSA Sparse MLA

[qiyueliuhuo23]

Motivation

Add an optional FlashMLA forward backend for GLM-5.2 DSA Sparse MLA.

The default PyTorch, TileLang, and cuDNN DSA behavior remains unchanged.
The FlashMLA runtime is maintained separately in the public repository:

https://github.com/DeepLink-org/flashmla-optimization/tree/precise-flashmla

The precise-flashmla branch aligns FlashMLA sparse-prefill output and LSE
with the TileLang reference on SM90/H200.

Changes

• Add flashmla to the SparseMLABackend protocol.
• Add lazy FlashMLA backend dispatch in xtuner.v1.ops.sparse_mla.
• Add runtime availability and SM90 checks.
• Reuse the existing TileLang indexer and backward implementation.
• Add a compiled unit test for FlashMLA forward plus TileLang backward.
• Add FlashMLA installation instructions to the English and Chinese
  installation documentation.

Runtime constraints

• Requires the public precise-flashmla branch.
• Requires SM90 or newer GPUs.
• H200/CUDA 12.8 users should set FLASH_MLA_DISABLE_SM100=1.
• Current FlashMLA sparse-prefill integration supports kv_group=1,
  value_dim=512, and top-k values divisible by 128.
• Existing backends and default behavior are unchanged.

Validation

• Python static compilation passed for the modified Python files.
• The remote H200 validation record contains 3/3 passing CI checks.
• 16K packing performance:
  • Median step time: 4.846% faster.
  • TGS: 5.086% higher.
  • Peak memory: reduced by 0.154 GiB.
• 32K packing performance:
  • Median step time: 4.650% faster.
  • TGS: 4.432% higher.
  • Peak memory: reduced by 6.875 GiB.

CI Comparison

Round	local_loss (before → after)	Δ	reduced_llm_loss (before → after)	Δ	grad_norm (before → after)	Δ	TGS (before → after)	Δ	Memory (before → after)	Δ
R1	6.613820 → 6.631019	+0.2600%	6.281574 → 6.274125	-0.1186%	10.159761 → 11.102958	+9.2836%	1660.282 → 1660.762	+0.0289%	116.354 → 116.358	+0.0034%
R2	5.883135 → 5.911429	+0.4809%	5.488452 → 5.508216	+0.3601%	7.448886 → 13.050797	+75.2047%	1660.282 → 1679.164	+1.1373%	116.354 → 116.358	+0.0034%
R3	6.093675 → 6.110326	+0.2732%	6.088449 → 6.093432	+0.0818%	10.159761 → 8.957117	-11.8373%	1660.282 → 1667.970	+0.4631%	116.354 → 116.358	+0.0034%

Performance Comparison

Packing Length	Median Step-Time (s)	P90 Step-Time (s)	P99 Step-Time (s)	TGS (tokens/s)	Memory (GiB)
16K	4.548683 / 4.328250	4.571737 / 4.351493	4.618507 / 4.387413	28,489.920 / 29,939.000	104.919 / 104.765
32K	19.688900 / 18.773400	20.814340 / 20.062140	21.862641 / 21.509888	13,099.526 / 13,680.117	133.956 / 127.081

Performance Difference

Packing Length	Median Step-Time	P90 Step-Time	P99 Step-Time	TGS	Memory
16K	-4.846%	-4.817%	-5.004%	+5.086%	-0.146% (-0.154 GiB)
32K	-4.650%	-3.614%	-1.613%	+4.432%	-5.132% (-6.875 GiB)