Add Debenedetti2022Light and support for timm

[dedeswim]

This PR adds support for timm models and the models from A Light Recipe to Train Robust Vision Transformers as an example of timm models.

Support for timm is integrated in the following way:

• We use timm's models registry mechanism to register the functions which generate the models
• For the models registered in this way, we must associate an URL to download the model from. This URL is ideally a file uploaded to a GitHub release of our repo. E.g. here. The URL can be pre-determined even if the release does not exist yet, as long as the release tag is already known. This removes the burden of uploading the models to Google Drive which can be unreliable.
• When we call load_model, the user must pass the model name used to register the model to timm's registry. Then, the function checks if the model name exists in the registry. If it exists, then it is loaded using timm.create_model (which also takes care of downloading the pre-trained checkpoint). If instead, it does not exist, it loads the model using the way which already exists.

The models I added will be uploaded as a release to the repo (which is already present as a draft here), so they will be downloadable once the PR is merged and the release is generated.

[fra31]

Looks great, thanks!