BaseOutputTransport: resolve media sender via type(self).MediaSender

[shivgargsquad]

What

Resolve the media sender class from the transport's type at both construction sites in set_transport_ready() (base_output.py), instead of hardcoding BaseOutputTransport.MediaSender:

self._media_senders[destination] = type(self).MediaSender(...)

This lets a transport subclass provide its own sender implementation by shadowing the inner class — the same extension pattern the class already uses for destinations (register_audio_destination), per-destination mixers, and delivery (write_audio_frame). It completes the per-destination sender model introduced in #1697: the sender becomes a per-transport extension point.

Compatibility

Fully backward compatible. type(self).MediaSender resolves to BaseOutputTransport.MediaSender unless a subclass deliberately shadows the inner class; no behavior change for any existing transport.

class MyOutputTransport(FastAPIWebsocketOutputTransport):
    class MediaSender(FastAPIWebsocketOutputTransport.MediaSender):
        async def _bot_started_speaking(self):
            await super()._bot_started_speaking()
            ...  # transport-specific behavior

Use case

Telephony platform (FastAPI websocket transport, multiple carriers) where transport-specific behavior genuinely lives in the sender layer: per-carrier audio pacing interacting with the sender's clock task, clearing buffered audio on programmatic (non-VAD) interruption without the bot-stopped-speaking lifecycle (same need as #3872), and observability on bot speech start/stop from inside the audio task loop. Today the only way to express these is post-construction mutation of _media_senders, which #3872's workaround shows is already happening in the wild.

Tests

Added TestBaseOutputTransportCustomMediaSender covering:

• default transports still get BaseOutputTransport.MediaSender
• a subclass-shadowed MediaSender is used for both the default sender and per-destination senders

Closes #5166