[Feature] Integrate fused moe kernel

[JamesBrianD]

Checklist

• If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sgl-jax/discussions/new/choose Otherwise, it will be closed.
• Please use English, otherwise it will be closed.

Motivation

The tpu-inference released a new fast moe kernel. We should integrate it. See Fused MOE Pallas Kernel.

Related resources

• fused Moe vllm-project/tpu-inference#973
• [FusedMoE] Support sub-channel quantization: FP4, FP8, INT8, ... vllm-project/tpu-inference#1158