Fix VAD tie-break to prefer latest max-silence candidate

## Summary When `use_max_poss_sil_at_max_speech=True`, `get_speech_timestamps` currently picks the first candidate when multiple silence candidates have the same duration. ## Problem `max(possible_ends, key=lambda x: x[1])` breaks ties by first occurrence, which biases the split point toward older silence segments instead of the most recent one near the max-speech boundary. ## Fix Use recency as a tie-breaker: - from: `key=lambda x: x[1]` - to: `key=lambda x: (x[1], x[0])` This preserves the longest-silence behavior and, on ties, chooses the latest candidate. ## Tests Added `tests/test_vad.py`: - `test_get_speech_timestamps_prefers_latest_silence_on_tie` Ran: - `pytest -q tests/test_vad.py` (Passed)