Qwen3-TTS PR 3: sentence chunking and error status for /v1/audio/speech

[merceod]

Long inputs to /v1/audio/speech are synthesized as ordered sentence chunks.

speech_chunking.split_sentences groups sentences (Latin and CJK terminators, paragraph breaks, over-long sentences wrap at clauses and then spaces) into pieces of about speech_chunk_max_chars. serving_speech submits one engine request per piece, keeps speech_chunk_lookahead pieces in flight while the current one streams, and concatenates the PCM in order for both streaming and non-streaming responses. Each autoregressive request stays within the length its model was trained for, the engine batches the pieces like independent requests, and a client seed advances per piece.

Opt-in per adapter (speech_chunk_min_chars, off by default) and overridable per request with sentence_chunking: true|false. Adapters that keep the default behave as before. Qwen3-TTS turns it on for inputs of 600+ characters.

Also in here: a streaming request now looks at its first engine result before answering 200, so a rejected request (bad voice, dead worker) comes back with its error status instead of an empty WAV, and a mid-stream error closes the stream.

Six commits, one per file. Tests in test/modular/test_speech_chunking.py cover the splitter and the handler on the stubbed router (ordered ids and playback, per-chunk seeds, streaming concatenation, opt-out, forced chunking, error status). Full test/modular passes.