Implement custom formatting in response_format

[sternakt]

Why are these changes needed?

These changes enable the developer to define custom structured output formatting of the response messages from the client.

Example use:

class Step(BaseModel):
    explanation: str
    output: str

class MathReasoning(BaseModel, FormatterProtocol):
    steps: List[Step]
    final_answer: str

    def format(self) -> str:
        steps_output = "\n".join(
            f"Step {i + 1}: {step.explanation}\n  Output: {step.output}"
            for i, step in enumerate(self.steps)
        )
        return f"{steps_output}\n\nFinal Answer: {self.final_answer}"

assistant = autogen.AssistantAgent(
    name="Math_solver",
    llm_config=llm_config,
    response_format=MathReasoning,
)

The assistant will output messages in the following format:

Step 1: [Explanation of the first step]
  Output: [Result of the first step]
Step 2: [Explanation of the second step]
  Output: [Result of the second step]
...

Final Answer: [The final answer provided by the reasoning]

For example, if the assistant is solving the square root of 16, the output could look like this:

Step 1: Break 16 into 4 x 4
  Output: 4
Step 2: Verify that 4 squared is 16
  Output: Correct

Final Answer: 4

This format is automatically generated by the format method of the MathReasoning model, ensuring clear and structured communication.

Related issue number

Closes #143

Checks

• I've included any doc changes needed for https://ag2ai.github.io/ag2/. See https://ag2ai.github.io/ag2/docs/Contribute#documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

[marklysze]

Thanks @sternakt, all tests passed. Will tweak the notebook text and then merge, thanks so much for this!

[sonichi]

@sternakt, one consideration is how we set the response_format to a Pydantic modeL when the config is loaded from a file (see loading OAI_CONFIG_LIST). E.g. is there a way to specify it in the config file or do we say that this isn't supported.
I personally don't think it's critical to have this supported when loading from a file but you may come up against it when you're working on it.

Loading the response_format from the config file is not supported at the moment, but that is a good point Right now, when I am loading OAI_CONFIG_LIST, I dynamically add the response_format to the configs

e.g.

config_list = autogen.config_list_from_json(
    "OAI_CONFIG_LIST",
    filter_dict={
        "model": ["gpt-4o", "gpt-4o-mini"],
    },
)

for config in config_list:
    config["response_format"] = MathReasoning

The generation of pydantic model from a json schema has been discussed before but they decided against implementing that if I remember correctly.

We could use something like datamodel-code-generator and generate a Pydantic datamodel when response_format is defined in the OAI_CONFIG_LIST

But I think we should address this in a separate FeatureRequest

The code can be simplified by setting llm_config={"response_format": ..., "config_list": config_list}