Implement RetryHandler on top of new internal API (closes #495)

[kiranandcode]

This PR reimplements RetryLLMHandler on top of the updated internal API.

When decoding fails, this handler:
1. Adds an error message to the conversation
2. Retries the completion with the error feedback
3. On success, returns the clean final message (without error messages)

The code is a bit long because call_assistant does several logical steps, but otherwise encapsulates all the relevant logic to a single place.

[eb8680]

There are at least a couple other points of failure that we'd want to cover with retrying logic:

• Server-side LLM API/network failures - this should be handleable within litellm (#494)
• Runtime tool call failure, i.e. a tool raises an unhandled exception - we'd need to separately handle call_tool and have it return an appropriate failure Message

[kiranandcode]

@eb8680 updated!

[eb8680]

We might also want to include the traceback in the feedback message. I think @datvo06 had done that at some point.

[eb8680]

Good to merge once tests pass. Nice work! This is much more robust and comprehensive than what we had before, and it should work really nicely in conjunction with the updated synthesis machinery #519 #520

[kiranandcode]

Awesome!