Start audio streaming for a call.

Initializes the streaming transcription handler for receiving audio chunks. The CallSession should already be created via POST /sessions/{id}/start-call.

This is used for production mock mode (audio file playback), microphone testing, and training mode (where only the AI persona's audio is streamed). Real telephony integrations (e.g., 3CX) call start_streaming_handler directly.

Args: session_id: Session ID for this call scenario: Optional scenario identifier (for orchestrator context) channels: Number of audio channels (1=mono, 2=stereo for speaker diarization). Default is 1 (mono). Use 2 only if audio has separate speaker channels. speaker: Optional fixed speaker label for a single (mono) stream. When set to "customer" or "agent", every transcribed turn is attributed to that speaker. Used by training mode, which streams only the AI persona's audio as the "customer". Only valid for mono (channels=1).