Fix backends in flash_attention and gemm

[xuzhao9]

To run PT2 cutlass backend, we have to add a cutlass submodule that has the same version as pytorch: https://github.com/pytorch/pytorch/tree/main/third_party

The version points to
https://github.com/NVIDIA/cutlass/tree/bbe579a9e3beb6ea6626d9227ec32d0dae119a49 which is 9 months old.
The FBGEMM cutlass is much newer.

Test plan:

$ python run.py --op gemm --mode fwd --only pt2_cutlass_matmul --num-inputs 1
      (M, N, K)    pt2_cutlass_matmul-speedup    pt2_cutlass_matmul-tflops    pt2_cutlass_matmul-gbps
---------------  ----------------------------  ---------------------------  -------------------------
(256, 256, 256)                                                    3.51871                    41.2349

Fixes #17

[facebook-github-bot]

@xuzhao9 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[FindHao]

LGTM

[facebook-github-bot]

@xuzhao9 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[facebook-github-bot]

@xuzhao9 merged this pull request in 17b38a4.