[Major] Add support for Mixtral8x7b

[cylinbao]

Add simulated quantization for Mixtral8x7b.
One major difference to Llama is that we move the activation quantization to after the gate operation of the SpareMoeBlock.
I also update the transformers library version to 3.39.0 for better support on the Mixtral model.
Currently, we have 4.41 perplexity on wikitext2 for W4A4 quantization.

[happierpig]

Note that the current implementation is shared reordering indices for all sparse experts. Ideally, different experts should have separate sampled and generated indices. However, due to the efficiency considerations (we want to fuse the reorder operator into the previous layer norm operator) and real evaluation (no large accuracy difference between these two ways), we choose shared indices.

[happierpig]

Note that the current implementation is shared reordering indices for all sparse experts. Ideally, different experts should have separate sampled and generated indices. However, due to the efficiency considerations (we want to fuse the reorder operator into the previous layer norm operator) and real evaluation (no large accuracy difference between these two ways), we choose shared indices.