Use Official PyTorch FA3 Builds

Hello Nanochat friends, The Pytorch team has released official versions of the FA3 builds here: https://download.pytorch.org/whl/flash-attn-3/. You can easily install these builds on a Hopper machine via pip install flash-attn-3 --find-links https://download.pytorch.org/whl/flash-attn-3/.

I recommend we transition away from the Huggingface Kernel (varunneal/flash-attention-3) in favor of an official release. (Historical note: this kernel was only necessary when the official branch of FA3 was not compatible with torch.compile.)

Quick troubleshooting tip: the Pytorch FA3 wheels seem to be built only on CUDA 12.6. As such, you will need to install the CUDA 12.6 toolkit

sudo apt install cuda-cudart-12-6

on your machine, if it is not installed already.

If this change is useful I will be happy to make corresponding PRs. Thank you!