litellm updates to support only messages llm kwarg

[dtam]

this pr:

• drops support for prompt and instructions and msg_history guard( and Guard( and Guard.from( arguments in favor of the streamlined messages argument that is now defacto for openai and other models
• adds reask_messages support, not supplying this and having reask enabled will fall back to a message composed of our default Prompt and Instruction prompts to keep legacy support
• drops special injection of json: xml: string: into llm input which we previously inferred and added
• drops supports for most of the legacy Callables in favor of litellm support except for hugging face pipelines and the arbitrary callable for local inference support
• updates message validation to be per message so that fix filter and refrain now work properly
• updates docs
• updates cookbook/example notebooks

needs
guardrails-ai/interfaces#13
guardrails-ai/guardrails-api#78
guardrails-ai/guardrails-api-client#26

[JosephCatrambone]

Pre-initializing lists in _init functions can have odd behaviors sometimes.

def foo(my_list=[]):
  my_list.append(len(my_list))
  return my_list

>>> foo()
[0]
>>> foo()
[0, 1]
>>> foo()
[0, 1, 2]

If we're modifying and appending to reask_messages internally, we should instead default this to None and check inside the method if it's None.