make triton attn req pre-mlri tagged triton

[vchiley]

I fork triton and rename it at triton_pre_mlri, triton diff here

llmfoundry/models/layers/flash_attn_triton.py is copy pasta from HazyResearch flash_attn_triton where I modify imports to be

import triton_pre_mlir as triton
import triton_pre_mlir.language as tl

This PR will be used as a springboard to test torch2 updt

[vchiley]

~~blocking composer pr (waiting for updt img): https://github.com/mosaicml/composer/pull/2229~~ merged

[vchiley]

This is integrated into https://github.com/mosaicml/llm-foundry/pull/149 closing in favor of that PR.