WIP: Eval checkpointing

[daanelson]

This adds a few features to fine-tuning:

• Saves checkpoints every n_epochs/steps
• Critically, patches the HF trainer s.t. deepspeed only saves the peft weights, instead of the entire 13/26 GB model every checkpoint.
• Returns best performing model according to eval loss
• Automatically splits train dataset into train & eval if no eval dataset is provided using train_test_split input

WIP b/c I still need to ensure the non-happy path (train_test_split=1) works.