This PR refactor our LLaMA lobe to become generic enough and version agnostic. I cannot share the associated recipe for now, but it works ok. I still may have a few question open w.r.t checkpointing, but I think that this should be ok. I think that the current state of our LLaMA lobe is a real problem as it may hide the capacity of SB to adapt and to enable people to develop their custom code.

@poonehmousavi Multiwoz will need to be slightly refactor around this code I think. As a rule of thumb here are the potential changes:

1. BNB and PeFT are not expected to be on the lobe. These are external things that must be managed at the yaml level.
2. SB has its own adapter scheme, we should rely on this as it's more transparent than PeFT. This does not prevent function like preparation of kbit training to be applied in the training recipe, but not in the lobe. We can also use PeFT adapter class with our adapter method.

To-do:

• refactor multiwoz
• dive into the tests that touch llama2.py and change them.