Voice mode: surface session and request errors

[stippi]

Polish for voice mode (#234), mostly around errors that were swallowed.

• ChatGPT subscription providers are not offered for voice. A provider with codex_auth has no access to the Realtime API; the settings no longer list it, and a voice.json naming one fails with that reason instead of "has no api_key".
• A rejected session fails with the server's reason. OpenAI accepts the WebSocket even with an invalid API key, then sends error and closes. connect() counted the handshake as success, so the agent reconnected forever and the status stayed on Connecting…. A connection now counts once session.created arrives.
• The session settings are confirmed. connect() sends session.update and waits for session.updated; a refused update (unknown voice or transcription model, a tool schema) fails the attempt instead of leaving the session on server defaults without tools or instructions.
• Errors are told apart by the client event they answer. response.create, response.cancel and conversation.item.truncate carry an event_id prefix. Errors for our cancel or truncate are logged only; one for our response.create frees the floor (before, any error did, even while a response was still coming). Every other error, and a response.done with status failed, shows as an Error line in the transcript and turns the status chip red (message in the tooltip) until an answer completes again.
• Create watchdog. The floor is given back when response.create gets no response.created within 10 s.
• Interrupted answers are marked in the transcript that seeds a renewed session, since their transcript says more than the user heard.

docs/voice-mode.md is updated.

Tests

• voice::tests: rejected session, close before the session started, rejected reconnects end voice mode instead of looping, rejected settings, a survived error shows until an answer completes (and a cancel error does not show), failed answer reason, interrupted mark in the renewal seed.
• voice::floor: create watchdog fires / is disarmed by response.created.
• llm::realtime: error causes from event ids, status_details reasons.
• voice::config: a codex_auth provider serves no realtime.

cargo fmt --check, cargo clippy -p llm -p code_assistant_core -p ui_gpui --all-targets -D warnings, cargo check --workspace --all-targets clean.