ScatterMoE to support LoRA Adapters

[fabianlim]

This issue to be done after #99

the kernel-hyperdrive has support for LoRA adapters, but need to install it properly into the plugin

• perhaps have the prepare happen during augmentation (still use the model_loader). in augmentation we intercept the peft_config
• ensure that checkpoint saving works correctly with LoRA. Maybe its easier to just modify the fsdp_save and load functions to support adapters.

In addition we defer the following things

• handle the mixed precision here, as we can now get the --bf16 or --fp16 flags easily.

(stretch) Furthemore, we could also improve the lora triton kernels to support lora_r < 16