[AWQ][MoE] Support AWQ MoE

[catswe]

Description

Enabled AWQ. Ran on v6e-8.

Changes:

To preserve the memory benefit of AWQ, I dequantized on the fly and added zero_point parameters. The changes from adding zero_point is similar in logic as scale parameters

tpu_inference/layers/vllm/quantization/awq.py

• Referenced vllm/model_executor/layers/quantization/awq_marlin.py
• Referenced tpu_inference/layers/vllm/quantization/fp8.py
• Referenced tpu_inference/layers/vllm/quantization/mxfp4.py
• Defined get_fused_moe_quant_config since otherwise attribute access error
• Implemented create_weights directly and uses extra_weight_attrs because upstream has no base AWQ MoE method to inherit from, unlike the other TPU MoE methods

Tests

Due to changes from #1659, doing pip install -e . may be needed before running the following command

MODEL_IMPL_TYPE=vllm vllm serve QuantTrio/MiniMax-M2-AWQ --gpu-memory-utilization=0.95 --tensor_parallel_size=8 --enable-expert-parallel --dtype bfloat16 --download-dir /dev/shm

Checklist

Before submitting this PR, please make sure:

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have made or will make corresponding changes to any relevant documentation.

[catswe]

i prefer creating a follow up pr!

[kyuyeunk]

i prefer creating a follow up pr!

thanks! will merge it!