Adding a train helper function and updating defaults of MinibatchAlgorithm and LLMJudge

[chinganc]

Adding a high level train helper function.

This PR also makes the default of

• MinibatchAlgorithm: ensure_improvement: bool = True. This has been found to ensure stability. This setting is used in some other recent algos like in GEPA and Adaflow.
• LLMJudge: use_formatted_response: bool = False. The current default formatting is not robust enough. I suggest turning it off by default and just returns the llm's response.

[allenanie]

Like this name! Learner and Guide

[allenanie]

LGTM

[allenanie]

There are some improvements to this PR feature/train, but it can be a separate PR:

LLM Judge will not silently fail (if it outputs score=0 and rejects a parameter update, due to lack of context)