refactor: separate sparse FlashMLA attention

[CUHKSZzxy]

Summary

• Separate dense MLA and sparse DSA FlashMLA execution into dedicated implementations.
• Keep shared KV-cache fill, flattening, and paged FlashMLA primitives in the dense base while moving sparse index mapping and kernels to sparse_mla.py.
• Select the sparse implementation from the CUDA attention builder and derive scheduler metadata from the selected implementation.

Validation

• Pre-commit checks passed.
• Focused FlashMLA tests passed (7 passed).
• GLM-5.2 FP8 TP8 with FP8 KV cache, MTP5, and symmetric-memory all-reduce completed 80/80 SWE-Smith requests without CUDA errors, OOMs, graph breaks, or recompilation warnings.

Assistance

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