V3 support for many tasks as possible in Huggingface, starting with "text-to-image"

With V2, the `HuggingFace` class allowed to do inference on many different models performing different tasks. Moving to V3 many of the available tasks are not functional anymore.

I would like to see them working again with the V3 `ModelBuilder` interface, and to be able to use a short code snippet to deploy a model.
Here's the complete list of the tasks we used to support:
```
"text-classification",
"token-classification",
"table-question-answering",
"question-answering",
"zero-shot-classification",
"translation",
"summarization",
"feature-extraction",
"text-generation",
"fill-mask",
"sentence-similarity",
"automatic-speech-recognition",
"text-to-image",
"text-to-speech",
"audio-to-audio",
"audio-classification",
"image-classification",
"object-detection",
"image-segmentation"
```

The most important tasks for us are "text-generation" (already working on V3) and "text-to-image" (currently broken). If you think you can make at least these two work it would be great.

An example of model that we would like to test is **black-forest-labs/FLUX.2-dev**.

Aucun fichier ni test n’est nommé. Commencez par lire le chemin V3 ModelBuilder qui prend déjà en charge text-generation, puis étudiez le déploiement du modèle FLUX.2-dev fourni pour text-to-image. Le travail sera considéré comme terminé lorsque text-generation et text-to-image fonctionneront au minimum via V3 avec un court extrait de déploiement, et que la prise en charge plus large des tâches listées aura été traitée dans la mesure du possible.

Rédigé par le modèle d'indexation à partir du texte de l'issue.