Sync with vLLM, add `apply_router_weight_on_input` arg to `fused_moe`

Module name has to correspond to repo name

Fix Llama expert layout

Add Llama MoE

Update for build.toml changes

Vendor `w8a8_block_fp8_matmul` and `per_token_group_quant_fp8`

Normalize some directory names