Add split-K heuristic for decode attention

[Aya-ZIbra]

Summary: This diff adds an automatic split-K size heuristic for the Blackwell FMHA decode kernel to optimize GPU utilization.

Added get_splitk_heuristic() that automatically computes optimal split-K size .

The heuristic ensures split sizes are multiples of TileN (256) and disables split-K when only 1 split would occur.

Performance Benchmarks show consistent 15-34% speedup over Triton split-K across all tested configurations:

• Average speedup: 1.24x
• Min speedup: 1.15x (Batch=16, SeqLen=32768)
• Max speedup: 1.34x (Batch=128, SeqLen=8192)

Reviewed By: jianyuh

Differential Revision: D89016012

[meta-codesync[bot]]

@Aya-ZIbra has exported this pull request. If you are a Meta employee, you can view the originating Diff in D89016012.

[meta-codesync[bot]]

This pull request has been merged in pytorch/FBGEMM@3086dd201373085b01da748f663285f98b1572c8.