Add Knowledge-Distillation

[fmassa]

This PR adds the implementation for both soft and hard distillation, as present in the paper.

We introduce a DistillationLoss, which wraps a criterion and a teacher_model and applies both the original criterion as well as the distillation loss from the teacher.

The distilled models have a different return type than the others:

• during training, we return both the class_token predictions, as well as the dist_token predictions
• during inference, we return a single Tensor as the average of both predictions, so that the rest of the inference pipeline stays the same

[TouvronHugo]

Good for me.

[fmassa]

Merging to unblock, will upload the weights in a follow-up PR