Make loaded peft adapters optionally trainable

[snarayan21]

Previously, loaded peft adapters (from HF) were not trainable. This PR adds the option to continue training peft adapters that are initialized from some existing adapter HF checkpoint.

Added unit tests to make sure adapter parameters are trainable when needed. this requires setting pretrained=True since pretrained_lora_id_or_path is only used in that case.

[dakinggg]

do you have any sense why the default is False? mostly just wondering if i'm missing something, cause I'm not sure why they would have the default as fals.

[snarayan21]

@dakinggg i think just because they assume that a saved adapter should be used directly for inference by default:
https://huggingface.co/docs/peft/en/package_reference/peft_model#peft.PeftModel.from_pretrained.is_trainable

Like they probably expect most people to specify the peft config if they want to train new adapters instead of instantiating from existing adapters