[Api inference] Split up tasks by modality

[NielsRogge]

This PR proposes to split up the inference API tasks by modality, similar to the docs of Transformers.

Before

After

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Wauplin]

We thought about doing so with @osanseviero when writing up these docs but to be honest I find the split to be more confusing than being really helpful. Typically Chat completion can be considered "NLP" or "Multimodel" depending of if you are working with an LLM or a VLM. Someone could also search "Automatic Speech Recognition" under "Audio" while it's rightfully under "Multimodal". Also, I feel that the most demanded tasks for Inference API (e.g. text to image, chat completion, image text to text) are in the same category (multimodal) which will eventually be peaking up on the rest. Keeping things simple with only 1 layer avoids having to think about these questions.

[julien-c]

yes i agree with @Wauplin

[NielsRogge]

Ok fair :) will close this one