Add python API for flash-attn

[YangKai0616]

This PR provides a template for a flash-attn Python API. Currently, based on a sycl-tla kernel, I have reproduced part of flash-attn’s functionality, such as fwd and varlen_fwd.

I see that the current implementation of the kernel does not expose an external interface, and the test files are quite limited in scope. If possible, perhaps we can jointly maintain a common API, similar to what Dao-AILab/flash-attention does. Subsequent unit tests can be based on test_flash_attn.py to remain consistent with the CUDA official interface.

During the reproduction I found some edge-case issues when running tests. As a result, I modified xe_flash_attn_prefill_epilogue.hpp and xe_flash_attn_prefill.hpp.

If you are interested, we can discuss this further. If there are any issues with the current code, please let me know. Thanks!

Current method to build the API:

cd /workspace/sycl-tla/examples/06_bmg_flash_attention/flash-attn
CUTLASS_SYCL_SRC_DIR=/workspace/sycl-tla pip install --no-build-isolation .

After that, you can run tests using the same import statements as in test_flash_attn.py.

[yao-matrix]

@rolandschulz, could you help review? Thx very much. The context is we are integrating flash-attention-2 kernel to Hugging Face, so need align API w/ CUDA to make it easy to integrate, thx,

[Antonyvance]

Need to figure out how to place python package, torch extension and associated triton kernel. Let's hold this since it requires a redesign.