Switch from fbgemm-gpu w8a8 scaled matmul to vLLM/marlin-kernels

[danieldk]

What does this PR do?

Performance and accuracy of these kernels are on par (tested with Llama 70B and 405B). Removes a dependency and resolves some stability issues we have been seeing.

Two test snapshots had to be changed. It seems like the perturbations come from switching from bfloat16 (which was the only dtype supported by fbgemm FP8 quantization) to float16 which is our default dtype for Llama (even though recent Llama checkpoints are bfloat16).

I added an environment variable USE_CUTLASS_W8A to force use of the cutlass kernels on compute capability 8. The question why we use FP8-Marlin on compute capability 8.9 comes up occasionally (because it's faster and probably more accurate because it's in FP16), so this is a backdoor to use W8A8 on 8.9 in case somebody really wants it.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[drbh]

LGTM