Parallelize build of 16 backward kernels

[danthe3rd]

Rebuilding the backward kernels now takes ~1:30mn (vs something like 8mn before). I believe there is more work done than before (total CPU time), but it's more convenient this way especially for development. We now have 35 files to build (not counting flashattention which is built separately).

Main changes:
(1) Split each kernel in its own file (so compilation is parallel)
(2) Only compile kernels corresponding to the current __CUDA_ARCH__ (it still takes ~30s to process files we don't compile due to includes and more)

[fmassa]

Awesome work @danthe3rd !

[fmassa]

This is awesome, thanks!

What do you think about not committing the autogenerated files to the repo, but instead running generate_kernels.sh inside setup.py ?

One drawback I see of doing this is that it will not work on Windows (although looks like some things don't compile properly on windows anyway, see #365 )

Another one is that if we don't commit the autogenerated files, things start to get a lot more "magical" (see e.g., all the code-gen things that PyTorch generate)

[danthe3rd]

It would totally make sense if we start generating more code, but as the number of files is quite limited, I believe it helps readability (and is simpler) to have the files already generated in the repo. But totally up to you