[Dhirenderchoudhary]

ModelType.AUDIO_GENERATION already lists /v1/audio/speech and /v1/audio/transcriptions, but the inference service didn't serve them, so speech models couldn't be used through InferiaLLM. This adds both, OpenAI-compatible, so a speech-to-text or text-to-speech deployment works with an API key (or the sandbox header) like a chat model.

• POST /v1/audio/transcriptions: multipart file + model (other form fields like language pass through). Returns the upstream JSON, e.g. {"text": "..."}.
• POST /v1/audio/speech: JSON {model, input, voice, ...}. Returns the audio bytes with the upstream content type.
• Both resolve deployments with model_type = audio_generation and go through the same rate limit, quota, usage log and circuit breaker as the other routes.
• New engine name speaches (speaches, an OpenAI-compatible server for faster-whisper and Kokoro/Piper). Any OpenAI-compatible engine works too, since the paths are on the base adapter.
• GatewayService.call_upstream_raw for these two, because call_upstream only does JSON in/JSON out. It reuses the same URL validation, header sanitizing, concurrency limit, size cap and breaker.
• The REST deployments list now returns model_type. It was already in the gRPC Deployment message but got dropped, so the dashboard couldn't tell an STT model from an LLM.

Testing:

• src/tests/inference/test_audio_endpoints.py: calls the real routes with an httpx.MockTransport upstream. It checks the multipart body reaches upstream with the deployment's inference_model, audio bytes and content type come back, missing fields are a 400, and upstream 4xx is passed through.
• test_listing_includes_model_type for the list change.
• Ran it locally in docker with speaches (faster-whisper-tiny.en, Kokoro-82M) registered as external deployments, through /inf/v1 with an API key: TTS returned a 24 kHz WAV in 2.7s, and sending that WAV to STT came back as the exact sentence. A browser webm/opus recording also transcribes correctly.
• Full suite with the CI ignore list: same failures as main.

[Priyanshu-u07]

Reviewed this and pushed a fix commit (fbd39a0) for three issues:

• Upload size: the transcription endpoint read the whole file into memory before the API key was checked, with no size limit, so a fake key plus a huge file could exhaust the inference service's memory. Uploads over 25 MB are now refused before reading.
• TTS input: no limit on input length; now capped at 4,096 characters (same as OpenAI), since TTS is slow and long texts block the speech server.
• Content types: the upstream's content type was passed straight through, and /inf shares the dashboard's origin. Only audio, JSON and text types are passed now, with nosniff.

Added 3 tests; they fail without the fix. Full suite shows the same failures as main.

Note for the canvas (#429): an LLM answer over 4,096 characters would now get a 400 from TTS, though voice replies are far below that.

Follow-ups, not blocking: audio requests don't count toward the token quota yet and behind nginx the default 1 MB body limit will reject real recordings.