Add multiclass support

[mwachnicki]

Summary of issues we have discussed today on the call:

• Divide classes into Classifier and Regressor
• Current BertClassifier class as a base for both
• Interface:
  • Classification: predict method returning classes and predict_scores method returning probabilities
  • Regression: predict method
• _predict_logits as a protected method that is used by predict/predict_scores to get meaningful results
• Refactor names of classes
• Make sure that the cross entropy loss uses softmax and is the best choice in terms of numerical stability, etc.
  • Actually, the best option would be to have the loss function as a parameter with the default value
• Similarly, there can be a possibility to choose the optimizer and its parameters

[MichalBrzozowski91]

Thanks for the review :) I added the main changes concerning division classes into Classifier and Regressor I did not implement the custom loss and optimizer:

• For me, it does violate the YAGNI principle - I am not sure this is needed now.
• If necessary, I would rather add it in another PR. This issue is separate from adding the multiclass support.

[mwachnicki]

I would also love to see the same basic tests for the Regressor classes.