Support audioStreamEnd in send_realtime for realtimeInput

[guptaishaan]

The Gemini Live API's send_realtime_input accepts an audio_stream_end parameter (a boolean flag) that forces the model to flush any buffered audio immediately when VAD (voice activity detection) is enabled. This is useful when the client wants to signal the definitive end of an audio stream without waiting for VAD to detect silence.

This change threads audio_stream_end support through all three relevant layers:

1. LiveRequest (live_request_queue.py): adds an audio_stream_end: Optional[bool] field alongside the existing activity_start/activity_end fields, and exposes a send_audio_stream_end() convenience method on LiveRequestQueue.
2. base_llm_flow.py: handles the new audio_stream_end field from LiveRequest by calling llm_connection.send_realtime(True).
3. GeminiLlmConnection.send_realtime (gemini_llm_connection.py): when the input is the boolean True, calls session.send_realtime_input(audio_stream_end=True) to forward the signal to the Gemini API.

A regression test verifies that calling send_realtime(True) correctly invokes send_realtime_input(audio_stream_end=True) on the underlying session.

Fixes#2887

[wuliang229]

Hi @guptaishaan. Thanks for the PR. We are reviewing #4490 and will go with that one.