OpenAI-compatible Custom-LLM relay for the voice agent

Called server-to-server by the ElevenLabs agent (Custom LLM) for each turn (#5399). Instead of running inference, it relays the turn over Socket.IO to the user's desktop, runs the local agent there, and streams the reply back as OpenAI chat.completion.chunk SSE. Authenticated by the shared ELEVENLABS_CONVAI_LLM_SECRET (presented as the Custom-LLM API key), not a user JWT. The user is identified by verifying the signed token the client set via startSession, read from user / user_id at the top level or from elevenlabs_extra_body — minted at get-signed-url and bound to the authenticated caller, so a raw user id cannot be spoofed here. A turn whose user message is a read-back request (the desktop pushing a deferred answer back into the live session to be spoken) is answered here from that message, without waking the desktop.