fix(copilot): ensure requests don't end on an assistant prefill

[arminanton]

Copilot's proxy rejects assistant-message prefill with "This model does not support assistant message prefill. The conversation must end with a user message." _ensure_copilot_request_ends_with_user appends a user turn on any path that would otherwise leave a trailing assistant turn, so reasoning/continuation requests stay valid on Copilot. Two call sites in the (streaming + non-streaming) interruptible API paths.

[teknium1]

Thanks for the Copilot compatibility investigation.

Automated hermes-sweeper review found that current main already provides this guarantee:

• agent/conversation_loop.py:1904-1946 records a truncated assistant response and then appends a user continuation before issuing the next request.
• agent/conversation_loop.py:907-917 removes thinking-only assistant prefills from the per-request wire copy, preserving role alternation without fabricating a request-only turn.
• tests/run_agent/test_run_agent.py:4768-4788 verifies that the continuation request ends with a user message.

Closing as implemented on main.