Fix CI after torchmetrics update

[danthe3rd]

Stack from ghstack (oldest at bottom):

• Refactor5: Generic way of registring operators for xformers.info #561
• Refactor4: Split fw/bw operators #560
• Refactor3: Embed flash's PY files as xformers._flash_attn #557
• Update FlashAttention #495
• Refactor1: Move files to fmha/ folder #555
• -> Fix CI after torchmetrics update #564

It now takes an argument: https://torchmetrics.readthedocs.io/en/stable/classification/accuracy.html
Somehow this is failing with a SEGFAULT on my A100 (in a triton kernel):

#0  0x00007fffc0f62e10 in ?? () from /lib/x86_64-linux-gnu/libcuda.so
#1  0x00007fffc0f9303c in ?? () from /lib/x86_64-linux-gnu/libcuda.so
#2  0x00007fffc0f2ea13 in ?? () from /lib/x86_64-linux-gnu/libcuda.so
#3  0x00007fffc0f94603 in ?? () from /lib/x86_64-linux-gnu/libcuda.so
#4  0x00007fffc119e4a0 in ?? () from /lib/x86_64-linux-gnu/libcuda.so
#5  0x00007fffc0f3728f in ?? () from /lib/x86_64-linux-gnu/libcuda.so
#6  0x00007fffc0f3999f in ?? () from /lib/x86_64-linux-gnu/libcuda.so
#7  0x00007fffc0fdb1c2 in ?? () from /lib/x86_64-linux-gnu/libcuda.so
#8  0x00007fff502234c0 in _launch ()
   from /data/home/XXXXX/.triton/cache/704a3e6949e60326bc68d18a620bee50/layer_norm_fw.so
#9  0x00007fff3c0eea25 in launch ()
   from /data/home/XXXXX/.triton/cache/2cebb5590a024a2e06fe9de08c6b7079/k_dropout_bw.so
#10 0x0000555555698422 in cfunction_call (func=0x7fff3c6e5760, args=<optimized out>, kwargs=<optimized out>)
    at /usr/local/src/conda/python-3.10.6/Objects/methodobject.c:552