[Chatllama] Add support to other open-source models

[diegofiori]

Description

LLaMA is a great model, however the weights license can be a huge barrier for users interested in the open-source LLM capabilities. The actual RLHF is designed around the LLaMA model, but it can be extended to other open-source models, like Flan-T5 or OPT.

TODO

• [x] Extend support of ChatLLaMA to HuggingFace models
• [ ] Add support for the encoder-decoder architectures (AutoModelsForSeq2Seq)
• [x] Add support for LLaMA from HF.
• [x] Add support for Cerebras GPT.
• [ ] Test the support for
  • [ ] FlanT5
  • [x] GPTJ
  • [x] GPTNeoX
  • [x] OPT
  • [x] BLOOM
  • [x] BLOOMZ
  • [ ] Galactica
• [ ] Implement unittest for the code