Add Support for PEFT fine-tuning

[Pierpaolo Sorbellini]

Description

Supports for PEFT for chatllama models and trainings

TODO

• Add PEFT to Enable Parameter efficient fine-tuning in actor, reward and critic models.
• Check RLHF stability.
• Check quality of adapted models vs. fully trained vs. reduced training requirements.
• Update README.

[Mark Schmidt]

Ideally 4bit PEFT like https://github.com/johnsmith0031/alpaca_lora_4bit, as this enables training 33b (30b) in under 24GB of VRAM on consumer cards like a single 3090.