Caching model responses for quicker development and branching of runs

[jas-ho]

Note: this needs to be defined more carefully.
Currently, this is mostly a bookmark for a discussion we had with @spirali and @gavento .

The rough idea would be to allow an actor to read responses from a context instead of querying an API.

How this would help?

• could speed up debugging by avoiding slow API calls
• could help debugging by providing a reproducible / deterministic actor response
• could allow branching experiments where, e.g., a run of a debate which looks promising up to round 10 can be repeated with several possible continuations after round 10