* Support a callback `on_llm_new_token` that users can implement when `OpenAI.streaming` is set to `True`