Consider using pre-built Flash Attention kernels via `kernels`

Nobody has claimed this yet.

Hey,

I am Sayak from the [Kernels](https://github.com/huggingface/kernels) team at Hugging Face. I noticed that this project uses Flash Attention which includes a long build time. We ship pre-built binaries (which provide bit-exact outputs as the upstream) and thereby, we make it easy to use.

Using FA3 on a supported machine is as easy as:

```py
# make sure `kernels` is installed: `pip install -U kernels`
from kernels import get_kernel

kernel_module = get_kernel("kernels-community/flash-attn3")
flash_attn_func = kernel_module.flash_attn_func

flash_attn_func(...)
```

Let us know if you'd be interested in this and and we'd be happy to provide a draft of how it would look in your repo.

Start by locating the project's current Flash Attention integration and build/dependency path, then review the proposed `kernels` entry point using `pip install -U kernels` and `get_kernel("kernels-community/flash-attn3")`. Done would require an agreed integration design, supported-machine behavior, and validation that outputs remain bit-exact without the existing long build.

Written by the indexing model from the issue text.