Create Audio Speech Stream

Streaming TTS — same model alias and request body as /v1/audio/speech, but the worker emits Server-Sent Events (audio.chunk, audio.timestamps, audio.done) so customers building voice agents can play audio chunks as they synthesize. The final audio.done event carries the assembled Bunny URL + usage; the gateway's _capture_sse_usage() reads the usage object from that event and bills the customer's account exactly like sync mode.

Workers that don't expose :stream will 404 here — that's correct: the catalog row's supported_endpoints advertises which models speak streaming, and customer code is expected to check capabilities before calling. Today: Inworld TTS Mini + Max, and Qwen Audio 3.0 TTS (whose upstream is natively duplex).