refactor alora train.py to work with newer versions of trl

[jakelorocco]

We are currently locked at trl==0.19.0 because of our alora code. We need to remove the data collator and change the parameters passed to the SFT Trainer.

Some resources:

• DataCollatorForCompletionOnlyLM huggingface/trl#3827
• https://huggingface.co/docs/trl/en/sft_trainer#trl.SFTConfig.completion_only_loss