fix: bound DSA prefill score memory

[CUHKSZzxy]

Depends on #4853. Supersedes #4892 by carrying its sparse FlashMLA refactor together with the DSA memory fix.

Summary

• Separate dense MLA and sparse DSA FlashMLA execution into dedicated implementations, with sparse implementation selection handled by the CUDA attention builder.
• Bound the FP32 DeepGEMM DSA prefill score tensor by chunking query rows under a configurable 512 MiB budget while flattening and reusing the indexer K cache once.
• Reserve the same DSA score workspace before automatic KV-cache sizing so cache allocation cannot consume the transient runtime budget.
• Keep score computation and top-k selection within consistent prefill and decode/fallback paths, and add coverage for chunking, result parity, and runtime-memory estimation.

Validation

• Pre-commit checks passed.
• Targeted DSA and executor tests passed (31 passed, 3 CUDA-dependent skipped).
• Focused FlashMLA tests passed (7 passed).
• Full GLM-5.2 LCB GPU reproduction is pending.

Assistance

Assisted with Codex + GPT-5.6-Sol xHigh, reviewed manually