ScatterMoE to support Quantized PEFT

[fabianlim]

The PR #99 includes the ScatterMoE module drop-in to enable expert parallel for mixture-of-expert models, but the class currently only supports full-finetuning and LoRA.

This issue is to extend it to also support quantized peft. Currently it is incompatible with the accelerated-peft plugin on many levels

• firstly the MoE models may have 3D tensors for experts, not sure if bitsandbytes support it.
• since we perform a complete model swap without caring about the base layers, we will ignore the quantized modules when doing the model swap

This requires a think through. The best outcome will be compatiblity with quantized_peft