fix(stt): preserve provider speech end timing

[swayamg20]

Summary

• add a provider-neutral wall-clock speech_end_time to SpeechEvent
• prefer that timestamp when STT-driven endpointing derives the last speaking time
• populate it for Deepgram Flux from final-word timing and re-anchor timestamps after reconnects

Why

Flux emits a final transcript followed by a bare END_OF_SPEECH event. Without an explicit VAD, that bare event previously replaced the speech-end timestamp with signal-arrival time, so min_delay used a different starting point depending on VAD configuration. Carrying the actual final-word time preserves the intended last-speaking-time semantics.

Testing

• make check (format, Ruff, and mypy across 632 source files)
• uv run pytest tests/test_agent_session.py --unit --no-concurrent -q (81 passed)
• uv run pytest tests/test_plugin_deepgram_stt.py --plugin deepgram --no-concurrent -q (14 passed)
• broader unit run: 1,845 passed and 5 skipped; 9 room-test setup errors require the unavailable local livekit-server binary

Closes #6726

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 1 additional finding.