fix(deepgram): end_time should be the last word, not the first

[mdylan2]

live_transcription_to_speech_data takes both start_time and end_time from the first word, since next() returns the first item of the generator:

start_time=next((word.get("start", 0) for word in alt["words"]), 0) + start_time_offset,
end_time=next((word.get("end", 0) for word in alt["words"]), 0) + start_time_offset,

prerecorded_transcription_to_speech_event in the same file uses alt["words"][-1]["end"], which is what this changes the streaming path to match.

It shows up on barge-in. audio_recognition.py drops a held transcript whose end_time falls before the cutoff, so "wait, I have a question" is judged on when "wait" ended and the whole turn can be discarded. Easy to miss because interim results are often a single word, where first and last are the same.

[devin-ai-integration]

✅ Devin Review: No Issues Found

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