[Inworld] Flush to drain decoder on every audio chunk from server

[ianbbqzy]

The emitter's non-PCM path still routes through _decode_task, which on FlushSegment does audio_decoder.end_input() + await decode_atask + audio_decoder = None. This means:

1. Without per-chunk flush(): All audio bytes accumulate in the decoder. The final output_emitter.flush() at line 1137 is the only FlushSegment. TTFB = time to receive all audio from server, not just the first chunk.

2. With per-chunk flush(): Each chunk triggers FlushSegment → decoder drains → _flush_frame() → SynthesizedAudio is emitted. TTFB = time to first chunk from server (correct).

This is an Inworld-specific issue because Inworld's HTTP API returns one JSON-line per audio chunk (each with its own WAV payload), unlike other providers which returns a single continuous raw PCM byte stream. For Inworld, each JSON line is a discrete audio segment, and delaying the flush until the end defeats the purpose of streaming.

====

Also updated timestamps parsing logic to always add a trailing space regardless if it's the end of a chunk or not. Because in Inworld case, end of a chunk could just be end of a phrase rather than end of a sentence

[ianbbqzy]

@tinalenguyen @davidzhao PTAL, Thanks!

[tinalenguyen]

Hi @ianbbqzy, thanks for the PR! Few notes and Q's:

• Could you update the TTS docstring with the new defaults?
• If each chunk is already flushed, might be worth removing the outer flush here

Also updated timestamps parsing logic to always add a trailing space regardless if it's the end of a chunk or not. Because in Inworld case, end of a chunk could just be end of a phrase rather than end of a sentence

Could you elaborate more on this? I also noticed that the last punctuation is not present, not sure if that was pre-existing behavior