Add an OpenAI TTS provider

[jason-shen]

Problem

[tts] provider currently accepts cartesia, deepgram, elevenlabs, minimax, speechify, vibevoice — but not OpenAI, even though github.com/sashabaranov/go-openai is already a direct dependency and most users running StreamCore already have an [openai] api_key set for the LLM. Someone who wants to try the project with exactly one API key currently can't: they need an OpenAI key for the LLM and a second vendor's key for speech.

OpenAI's /v1/audio/speech also supports streamed chunked audio, so it fits the existing SynthesizeStream contract rather than degrading to the one-shot path.

Proposed change

Add internal/tts/openai.go implementing tts.Client:

type Client interface {
	Synthesize(ctx context.Context, text string) ([]byte, error)
	SynthesizeStream(ctx context.Context, text string) (<-chan StreamChunk, error)
}

internal/tts/deepgram.go and internal/tts/speechify.go are the closest models to copy — both are HTTP providers returning PCM. Read internal/tts/http.go first; the shared request/stream plumbing lives there and a new provider should reuse it rather than hand-rolling a client.

Things the implementation has to get right, all of which the existing providers already demonstrate:

• Output format. The pipeline expects linear16 PCM, 16 kHz mono. Request pcm from the API and confirm the sample rate — if OpenAI returns 24 kHz you must resample; internal/tts/resample_test.go shows the helper that already exists for exactly this.
• Streaming. Emit StreamChunks as bytes arrive so playback starts before synthesis finishes. Close the channel exactly once, including on the error path.
• Cancellation. Barge-in cancels the context mid-utterance. Aborting must not leak the response body or the goroutine.

Wiring:

1. internal/config/config.go — reuse [openai] api_key; add tts_model and tts_voice to OpenAIConfig (models gpt-4o-mini-tts / tts-1 / tts-1-hd, voices alloy, nova, …). Default the model and voice in Load().
2. internal/tts/client.go — add case "openai": to newProviderClient, with the same ErrMissingAPIKey guard the others use, and add openai to the "supported:" list in the default branch's error string.
3. config.toml.example + README provider table.

Acceptance criteria

• [tts] provider = "openai" synthesizes speech end to end on a call.
• Streaming emits multiple chunks for a long utterance, not one buffered blob.
• Cancelling the context mid-stream closes the channel and releases the connection.
• Audio is 16 kHz mono linear16 by the time it reaches the pipeline.
• Unit test in the style of internal/tts/http_test.go using an httptest server.

Note for other providers

The same shape applies to Rime, PlayHT, Hume, Azure, and Google TTS. If you'd rather add one of those, that's welcome — open a separate issue referencing this one so two people don't land the same provider twice.

Pointers

• internal/tts/client.go — the Client interface and newProviderClient switch
• internal/tts/http.go, internal/tts/deepgram.go, internal/tts/speechify.go — closest existing patterns
• internal/tts/errors.go — ErrMissingAPIKey