Update autotune configuration to avoid crash on AMD devices

fix: adjust typo

feat: add tag for hfjob build

Make `RMSNorm` layer pure, rename to `LlamaRMSNorm`

feat: update to include rev in kernel for reproducible symbols

Build

black

Use `torch.amp` rather than deprecated `torch.cuda.amp`

Move `RMSNorm` to `layers`

Build

Add Triton-based layer norm from flash-attention