gemlite integration in torchao

[HDCharles]

Summary:

1. integrated gemlite kernels from https://github.com/mobiusml/gemlite
2. updated kernel wrappers to work with compile (needed to change acc_dtype to torch.dtype equivalent and added basic op/fake_op registration)
3. added bs>1 support to llama benchmark

on A100 results are not great for the kernel, likely needs further tuning before further integration makes sense.

Test Plan: benchmarks.sh

Reviewers:

Subscribers:

Tasks:

Tags: