port m train to use peft version of aloras

[jakelorocco]

When the m train code was written, peft had not yet added alora support directly. As a result, we are using a fork of peft for our alora support. This forked version is different from the alora support that eventually got added to peft.

Class names and parameter names / values changed between the versions. The PR that added the alora support directly to peft has an example of a fine-tuning script. We should make sure our fine-tuning example complies with that version.

Because of different parameter names / values, I believe aloras trained with this script will also stop working. For example, the original granite aloras do not work with the new peft version:

/opt/homebrew/Caskroom/miniforge/base/envs/gcom/lib/python3.12/site-packages/peft/config.py:165: UserWarning: Unexpected keyword arguments ['alora_invocation_tokens'] for class LoraConfig, these are ignored. This probably means that you're loading a configuration file that was saved using a higher version of the library and additional parameters have been introduced since. It is highly recommended to upgrade the PEFT version before continuing (e.g. by running `pip install -U peft`).

[jakelorocco]

fyi @nrfulton, I believe this will cause aloras trained with the script to not work with our huggingface backend once my rag-intrinsic changes are merged.

[nrfulton]

Assigning @avinash2692 for now, but we can pull in someone else if he doesn't have bandwidth.