The client-to-cloud WebRTC connection was being created with no ICE/TURN servers, causing it to fail ICE negotiation and time out after 30s.
Two bugs in CloudWebRTCClient.connect():

1. Double data extraction — webrtc_get_ice_servers() already unwraps response["data"] and returns {"iceServers": [...]}. But the client then called .get("data", {}).get("iceServers", []), looking for a non-existent nested "data" key. Result: empty list, TURN servers silently dropped.
2. Wrong config type — Even if the servers were extracted, a raw dict was passed to RTCPeerConnection() instead of a proper RTCConfiguration with RTCIceServer objects (which is what the local WebRTC path does correctly via create_rtc_config()).

The fix reuses the existing credentials_to_rtc_ice_servers() helper from webrtc.py to properly parse the ICE server response into RTCIceServer objects, and wraps them in an RTCConfiguration — matching how the local WebRTC path works.