feat: lora for accelerated moe

[willmj]

Add peft configs to fast moe augmentation so lora config is passed into prepare_scattermoe.
Restrictions:

• lora_config.r must be >= 16
• Must be using FSDP, since the scatteredExperts weights are not supported by peft's LoRA tuning, the overwritten FSDP save and load functions must be utilized here.
• Loading from a default lora adapter config may not work here, intended use-case is to run tuning from base model.

[willmj]

LoRA SD splits input and output linear layers into LoRA A and LoRA B layers as shown below in the logs:

scattermoe-lora.log

Compared to FT logs:

scattermoe-ft.log

The state dicts at the bottom are the ones right before they are saved to the converted checkpoint file after being run through the recover_original_state_dict_from_checkpoint function