feat: add multiple classes conditionning

[Bycob]

@beniz This PR is not ready, I have several questions about the integration of multiple class in joligen:

• The cls_semantic_nclasses seems to be used only by GAN models, for diffusion, class conditioning uses the f_s_semantic_nclasses parameter. Should we use cls_semantic_nclasses instead? cc @royale
• Is it fine to have multiple heads to embed the class, like I did? I'm not satisfied of this solution because it adds complexity to the code, I kept single head for backward compatibility with old models
• Check the backward compatibility with train_config, since *_semantic_nclasses is now an array instead of an int
• (As @royale suggested, we could sum embeddings of different classes instead of concatening them)