[DO NOT MERGE] Vinayak/moe final hashem

[carlushuang]

This PR combined vinayak/moe_final with PR:#126, kernel from hashem
Note: to enable triton optimization, need download triton code:

export VLLM_USE_ROCM_CUSTOM_PAGED_ATTN=1
export HIP_FORCE_DEV_KERNARG=1
export VLLM_USE_TRITON_FLASH_ATTN=false

export VLLM_MOE_PADDING=128
export TRITON_HIP_USE_NEW_STREAM_PIPELINE=1
export VLLM_MOE_MFMASWIZZLE=1 
export VLLM_MOE_MFMASWIZZLE_M_THRSHLD=128
# if set VLLM_MOE_MFMASWIZZLE_M_THRSHLD=0, this is just doug + vinayak
export VLLM_SCHED_PREFILL_COUNT=6

you can use benchmark in benchmarks/kernels/benchmark_mixtral_moe_dec.py to do benchmark

# test hashem's kernel with tp-2
VLLM_MOE_MFMASWIZZLE_M_THRSHLD=128  python3 benchmarks/kernels/benchmark_mixtral_moe_dec.py  --tp-size 2
 
# test hashem's kernel with tp-1
VLLM_MOE_MFMASWIZZLE_M_THRSHLD=128  python3 benchmarks/kernels/benchmark_mixtral_moe_dec.py  --tp-size 1
 
# test vinayak's triton kernel with tp-2
VLLM_MOE_MFMASWIZZLE_M_THRSHLD=0  python3 benchmarks/kernels/benchmark_mixtral_moe_dec.py  --tp-size 2
 
# test vinayak's triton kernel with tp-1
VLLM_MOE_MFMASWIZZLE_M_THRSHLD=0  python3 benchmarks/kernels/benchmark_mixtral_moe_dec.py  --tp-size 1