fix: reject history-rewriting handoffs with server-managed conversation

[sztoplover-bit]

Summary

• reject handoff input filtering when using server-managed conversation
• reject handoff history nesting when using server-managed conversation
• add regression tests covering conversation_id, previous_response_id, and auto_previous_response_id

Why

Fixes #2151.

When conversation_id, previous_response_id, or auto_previous_response_id is enabled, the server owns conversation history. Handoff input filters and nested handoff history rewrite the local history, which makes the next request diverge from the server-managed transcript.

Instead of allowing that mismatch to surface later as a confusing API error or incorrect behavior, this change fails fast with a clear UserError.

Testing

• .venv/bin/pytest tests/test_agent_runner.py -k 'server_managed_conversation'
• .venv/bin/pytest tests/test_run_step_processing.py -k 'handoff_can or missing_handoff_fails or multiple_handoffs_doesnt_error'

[seratch]

Thanks for suggesting this change. We've been receiving this type of fail-fast validation for this issue, but I don't believe that could resolve the issue (I agree it could mitigate the issue by preventing data issue tho). I came up with an alternative #2747; if you have any feedback, it'd be appreciated