refactor: clean up agent.step()

[cpacker]

Initial stab at cleaning up the agent step by removing dead code + turning the response into a response model instead of a tuple.

Follow-up PRs should do the following cleanup:

• Unify the type going into step (probably should drop support for str, make it None or Message?)
• Move unnecessary kwargs into configs
• Strip out ms (replace with session)
• Move input validation out of step and into the loop handler
  • This ensures that step corresponds to a single LLM call (unless a summarization happens, in which case we probably want to return a dual usage? or put a summarization event into the response model?)
• Potentially change the input of step to be a model as well (probably cleaner if we don't do this)

[sarahwooders]

lgtm!