Tighten BreezeTTS streaming path

[0xShug0]

Summary

This tightens the BreezeTTS streaming implementation around the PocketTTS-style ownership boundary. The session now only drives stream events, while the generator owns AR streaming state and returns decoded audio events to the session. The speech decoder owns the incremental codec decode state with left context and pending frames.

Breeze streaming now uses the incremental codec-frame path by default instead of the old one-output-per-text-chunk path. stream_frames_per_event controls event granularity, and stream_lookahead_margin controls how many trailing codec frames are held before emission to reduce boundary artifacts.

The offline generation path is unchanged.

Validation

Case	Result
Build	audiocpp_cli and audiocpp_server build cleanly
Default streaming	Passed, emitted 4 incremental audio events plus merged output without any opt-in subchunk flag
Offline A/B voice clone	Passed, artifacts/text match; timing 4594.314 ms vs 4452.668 ms (+3.18%)
Offline A/B voice design	Passed, artifacts/text match; timing 3375.125 ms vs 3375.804 ms (-0.02%)
Path-test compare summary	2 cases compared, 0 differences

Server TTFT

stream_frames_per_event	Server TTFT	Client first delta	Events	Wall
32	650.675 ms	663.389 ms	4	1844.125 ms
16	375.147 ms	375.717 ms	7	1836.698 ms
8	377.302 ms	377.861 ms	12	2062.941 ms

[0xShug0]

@Th-Underscore Credits to your PR. If you like, I can build on your PR directly instead of opening a separate PR.

[Th-Underscore]

Looks good! No need to rebase on my PR, this is perfect as is. I was thinking of a full Moshi/Mimi StreamingConv1d+StreamingConvTranspose1d port, but this is much better.

A small thing I didn't include myself but may be useful, though, is in BreezeTTSSession::next_stream_event() (src/models/breeze_tts/session.cpp) the
next_stream_audio() AR step can throw, which leaves stream_chunk_active_ true (and the
generator's stream set). A try-catch that calls end_stream() would make it work cleanly.

[0xShug0]

Looks good! No need to rebase on my PR, this is perfect as is. I was thinking of a full Moshi/Mimi StreamingConv1d+StreamingConvTranspose1d port, but this is much better.

A small thing I didn't include myself but may be useful, though, is in BreezeTTSSession::next_stream_event() (src/models/breeze_tts/session.cpp) the next_stream_audio() AR step can throw, which leaves stream_chunk_active_ true (and the generator's stream set). A try-catch that calls end_stream() would make it work cleanly.

Will do later!