[Kernel] add triton fused moe kernel for gptq/awq

[jinzhen-lin]

The current only option for using moe+gptq/awq is the Marlin kernel, but for the Marlin kernel, a single marlin_gemm_moe would launching num_experts CUDA kernels at least, while the fused_moe triton kernel only needs to launch one cuda kernel. This makes the Marlin kernel significantly slower than the fused_moe triton kernel.

This PR adds support for fused_moe triton kernel with gptq/awq.

Generation speed of deepseek-v3-awq (8*A100-SXM4-80GB, bs=1, short prompt)

	marlin moe kernel	triton fused moe kernel
w/o cuda graph	5.4tok/s	10.0tok/s
w/ cuda graph	10.5tok/s	25.3 tok/s

Note:

1. to enable cuda graph, use this pr [Bugfix] make moe_align_block_size compliable with cuda graph #12036
2. to enable this kernel

python -m vllm.entrypoints.openai.api_server \
    --served-model-name model \
    --model cognitivecomputations/DeepSeek-V3-AWQ \
    --tensor-parallel-size 8 \
    --trust-remote-code \
    --max-model-len 24576 \
    --dtype half \
    --max-num-seqs 16 \
    --gpu-memory-utilization 0.96 \
    --quantization moe_quant_int

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀