Add text-to-speech and text-to-audio task helpers for WaveSpeed provider

[chengzeyi]

What

Adds text-to-speech and text-to-audio task support for the wavespeed inference provider:

• WavespeedAITextToSpeechTask — sends the input text as text (the field WaveSpeed TTS endpoints such as wavespeed-ai/vibevoice and chatterbox/text-to-speech expect), mirroring how the fal.ai TTS helper handles the same translation.
• WavespeedAITextToAudioTask — reuses the base payload (prompt = inputs), matching WaveSpeed music/audio endpoints such as wavespeed-ai/ace-step/prompt-to-audio.

Both reuse the existing WaveSpeed submit-and-poll response handling already used by the image/video helpers.

Fix in textToSpeech()

textToSpeech() called providerHelper.getResponse(res, undefined, undefined, ...) without url/headers. That works for providers that return audio synchronously (fal, replicate), but WaveSpeed's helper polls an async task result and requires them. This PR passes url/headers the same way textToAudio() already does; providers that ignore these arguments are unaffected.

Why

WaveSpeed serves several audio models whose HF Hub counterparts currently have no inference provider at all (e.g. microsoft/VibeVoice-1.5B, Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice, k2-fsa/OmniVoice, ACE-Step/Ace-Step1.5, MiniMaxAI/MiniMax-Music3). We'd like to register those model mappings, which requires client support for the tasks first.

Testing

• pnpm run check, pnpm run lint, pnpm run build pass in packages/inference.
• Added textToSpeech / textToAudio cases to the WaveSpeed block in InferenceClient.spec.ts.
• Verified live against the WaveSpeed production API through the new helpers: textToSpeech (VibeVoice) returned an audio/mpeg blob, textToAudio (ACE-Step prompt-to-audio) returned an audio/wav blob.

I'm the maintainer on the WaveSpeed side (we operate the wavespeed provider). cc @Wauplin @SBrandeis @hanouticelina

🤖 Generated with Claude Code

https://claude.ai/code/session_0139Zu5TfHzGVyPeatY9HJAR

---

Note

Low Risk
Scoped provider additions plus aligning textToSpeech with an existing textToAudio pattern; no auth or data-model changes.

Overview
Adds WaveSpeed support for text-to-speech and text-to-audio, wiring new task helpers into the provider registry and reusing the existing submit-and-poll Blob flow used for other WaveSpeed media tasks.

WavespeedAITextToSpeechTask maps client inputs to the API text field (for endpoints like VibeVoice); WavespeedAITextToAudioTask keeps the default prompt-style payload for music/prompt-to-audio routes. Integration tests cover VibeVoice TTS and ACE-Step text-to-audio.

textToSpeech() now passes request URL and headers into getResponse, matching textToAudio(), so async providers (including WaveSpeed) can poll task status; synchronous providers that ignore those args stay unchanged.

^{Reviewed by Cursor Bugbot for commit 7e94f0c. Bugbot is set up for automated code reviews on this repo. Configure here.}