feat: patch missing inner thoughts on new openai models

[cpacker]

TODOs:

Refactor code to convert to inner thought kwarg format

• Add utility method inside of data types to convert Message objects
  • See note*
  • Add utility to convert ToolCalls into versions that have a kwarg included (add_inner_thoughts_to_tool_call)
  • Add flag put_inner_thoughts_in_kwargs to Message.to_openai_dict which will call utility to make put self.text (the content / inner thoughts) into ToolCalls via the utility
• Add utility method to convert functions list / tool objects (will require dumping and reading json strings of arguments)
  • add_inner_thoughts_to_functions
• Add utility method to convert response from OpenAI response to strip inner thoughts from kwargs and into the content field for data storage
  • unpack_inner_thoughts_from_kwargs
• Add flag to CLI to allow turning this on manually
  • I added a --no-content flag (yes = put inner thoughts in kwargs, no = leave as normal (gpt4 style), default = let us decide, eg for gpt4o turn it on, for gpt4 turn it off)
• TODO future PR - add support for our other API backends

*note: I was going to create a Message -> Message utility function that puts self.text into self.tool_calls (as an alternative to putting the logic for extracting inner thoughts from self.text into the function/tool call in the to_*_dict() methods), however upon further thought I think it's better to leave this logic inside the _to_dict() methods because it might vary significantly from provider to provider. Example of what I was going to do:

# TODO have this use model_copy
def put_inner_thoughts_in_kwargs(message: Message) -> Message:
    """Convert a message that has inner monologue in self.text to one that has it inside a tool call"""
    # message_copy = message.model_copy()
    message_copy = copy.deepcopy(message)

    # TODO
    # step 1: assert the role is assistant and tool_calls > 1
    # step 2: copy self.text into self.tool_calls[0], and make self.text = None
    # step 3: return the message copy

Please describe the purpose of this pull request.

Fixes issues with missing content / inner thoughts on newer OpenAI models like 4o-mini

How to test

memgpt run with gpt-4o-mini

Have you tested this PR?

Yes, gpt4o-mini works fine now:

Related issues or PRs

Closes #1555

[sarahwooders]

lgtm!