Flatten DefaultClassifier interface

[alanakbik]

This PR attempts to flatten the DefaultClassifier interface. We remove the distinction between tensor and non-tensor operations, temporarily removing JIT support until we find a more comprehensive solution.

Changes:

• The DefaultClassifier now has an embeddings argument in the init method, since all current implementations use embeddings. This removes the need for the _inner_embeddings property which no longer needs to be implemented in the subclasses.

• All implementing classes now have their embeddings argument called embeddings (rather than word_embeddings/document_embeddings etc.). Hopefully this introduces a bit more consistency across model classes.

• Previously, the get_prediction_data_points was called twice in the forward pass: once in forward_loss, then again in _prepare_tensors. It is now only called once, directly at the beginning of the forward_loss.

• Renames _embed_prediction_data_point to _get_embedding_for_data_point (in most cases, the embedding already exists)

• Changes _get_prediction_data_points to _get_data_points_from_sentence, changing the scope to extracting points for a since Sentence rather than the full batch

Additional changes:

• Removes the expectation maximization approach from the PrototypicalDecoder as it never really worked and required a special function in the model
• DataPoint now has __len__