[fa3] Upgrade fa3 on B200

[xuzhao9]

To run flash_attention on B200, we need to upgrade the submodule.

flash_attn_v3 does not support B200, but flash_attn_v2 support is done on B200.

Test plan:

python run.py --op flash_attention --metrics tflops --only flash_v2 --batch 8 --n-heads 16 --d-head 128 

  (Batch, Heads, SeqLen, Dhead)    flash_v2-tflops
-------------------------------  -----------------
              (8, 16, 128, 128)            60.5676
              (8, 16, 256, 128)           147.492
              (8, 16, 512, 128)           245.259
             (8, 16, 1024, 128)           315.111
             (8, 16, 2048, 128)           352.104
             (8, 16, 4096, 128)           353.912
             (8, 16, 8192, 128)           361.591
            (8, 16, 16384, 128)           330.587
                        average           270.828

[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 79c5302.