Polish current multiturn tests to catch regression

### What are you trying to build?

Check on and make sure the coverage of our existing multi-turn tests suffice

### Where are the current docs or utilities unclear?

Test coverage insufficient

### Expected improvement

Multiturn tests should cover all realistic cases, ie but not limited to
- Back to back turns w an early EOS break (race condition)
- Long conversation verifying prefix cache reuse across turns

### Additional context

_No response_