Add support for dbrx moe

[charlifu]

Dbrx-instruct model does not use nn.Linear to implement the Moe layer, which leads to AutoFP8 not able to quantize the moe weight. This PR adds quantization support for dbrx model's moe layer.

• Add a FP8DbrxExpertGLU module def to replace original DbrxExpertGLU model to quantize the weight and activation (if needed).
• Add a example to quantize dbrx model.
• PPL test using wikitext2: 3.91(fp16) vs 3.99 (fp8)