[Feature] Unsloth Studio APIs for audio, image, and video generation

[lijamez]

Unsloth Studio supports audio, video, and image generation over the web UI, but not programmatically.
This is a feature request to introduce APIs for audio, video, and image generation. When a model is requested, Unsloth Studio should, if necessary, unload the current model (after ongoing requests are satisfied) and load the API-requested model. This feature will be useful for agents like Hermes to generate non-textual content.

[mahiatlinux]

Most of this already exists. All three are reachable with an API key today (Authorization: Bearer sk-unsloth-...):

• image: POST /v1/images/generations, OpenAI-compatible. POST /api/inference/images/generate for the rest (LoRAs, ControlNet, img2img, inpaint, batches).
• audio: POST /v1/audio/speech (TTS) and POST /v1/audio/transcriptions (STT).
• video: POST /api/inference/video/generate. It starts a job and returns at once, then poll GET /api/inference/video/generate-progress.

Two real gaps behind the request:

1. No auto-load. /v1/images/generations returns 503 when no image model is loaded and the media routes return 409, so a model still has to be picked in the UI first. openai_api_auto_switch_model covers chat GGUFs only, and media_auto_unload_idle_seconds unloads but never loads.
2. Video has no OpenAI-shaped /v1 route, unlike image and audio.

I am working on (1): a media_api_auto_switch_model setting, off by default, that resolves model against the downloaded image and video models, waits for in-flight generations to drain, then runs the same load the picker would. POST /api/inference/video/generate gains an optional model for the same purpose. Nothing there starts a download, so an unknown name is refused rather than answered by whichever model happens to be resident.

(2) is not covered by that change.