Fix truncated Opus tail in streaming writer

[slime-turing]

Summary

• fix truncated Opus tails in the in-memory streaming writer
• keep Opus on a 48 kHz codec clock while resampling 24 kHz PCM input
• document why the final Opus/Ogg bytes must be read only after muxer finalization

Root cause

The in-memory Opus path was reading the output buffer before the muxer finished flushing on container close. That left the final Ogg/Opus pages out of the last response chunk, which shortened clips and could clip the final word.

Testing

• PYTHONPATH=$PWD .venv/bin/python -m pytest api/tests/test_audio_service.py -q
• PYTHONPATH=$PWD .venv/bin/python -m pytest api/tests/test_openai_endpoints.py -q -k "speech_endpoint or speech_streaming or speech_pcm_streaming"
• direct writer-level Opus encode/decode validation confirming ~2.089s output from ~2.09s input after the fix

[slime-turing]

Closing this for now because it overlaps the same root-cause fix already proposed in #448. I’m keeping the local branch and extra regression/docs work on my side until upstream merges #448 or another equivalent fix, then I can re-check whether anything unique still needs to be proposed.