[BUG] Quantization support for MoE models

[WoosukKwon]

After #2453 and #2542, the new fused MoE kernel used in Mixtral and DeepSeek-MoE does not support quantization.

[AlpinDale]

An easy, but hacky, solution would be to fall back to the old MoE code if quantized linear methods are detected. Unquantized models can use the new kernels.

[casper-hansen]

I was looking to import the MoE Triton kernel to AutoAWQ.

I was looking at two solutions:

1. Dequantize before using Triton kernel. Requires some implementation details since nn.Parameter is used with multiple experts. Unsure if you can loop over the experts and if it would be slow.
2. Implement dequantization in the MoE Triton kernel from @WoosukKwon's qmm branch.

Both introduce some complexity. I would prefer the second one for performance reasons though.