Small fixes for easier reproducibility

[janfrancu]

I've made a few fixes, that I've encountered while trying to reproduce the results from your article.

• fixed unbound local variable and logic around label/dataset balancing, otherwise the validation set was empty
• fixed requirements version specification for pytorch
• added a check for the existence of root directory for RAISE, as the dataset_creation/simple_dataset_generation.py is completely fine, when one of the real datasets is not properly loaded (should be probably extended to others as well)
• fixed openai/clip model factory registration, in order to support the configs in form RN50_tune, ViT-L/14_probe and other models available from the openai/clip library