support custom triton sdpa kernel in CUDA backend

[Gasoonjia]

🚀 The feature, motivation and pitch

Currnently one of the main gap for cuda backend is we don't support sdpa kernel in one step, but need to decompose it which introduces extra perf latency.
We should have a single triton sdpa kernel for CUDA backend.

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

[Gasoonjia]

We've got some init number on A100 and 5080, both beats the baseline (decompose sdpa kernel we are using). Working on trying new version kernel, and a consolidate prototype

[Gasoonjia]

PRs are landed / landing:
#15861
#15744