Bump kernels-community/flash-attn2 to v4 (#50009)
* Bump kernels-community/flash-attn2 to v4
The v3 branch of the flash-attn2 kernel repo ships an XPU build that predates
huggingface/kernels#921, so it has ahead-of-time compilation silently disabled
and is built against a sycl-tla revision with a flash-attention forward
regression. v4 carries both fixes.
v3 and v4 produce bit-identical logits, so this is a build-level change only.
* only on recent versions
---------
Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>