Get sparse feature sizes

[jupyterjazz]

In order to dynamically change sparse layers, we should be able to split a combined feature input into features that come from separate featurizers. The reason behind this is that new features will be appended to the corresponding splits and not to the combined representation. That's why we need to know the size of each sparse feature.

feature_sizes collection will have a following structure: Dict[Text, Dict[Text, List[int]]]
It will be a dictionary of attributes that have sparse features(e.g. text, response… we are excluding label at this point) to the dictionary of different kinds of features(e.g. sequence/sentence) to the list of feature sizes of different featurizers.

The collection should be created in the DIETClassifier class during the _create_model_data call and it should be assigned to the DIETClassifier class as an attribute.

feature_sizes should be persisted after training and loaded during fine-tuning. We will need to compare persisted feature_sizes to the current feature_sizes to find out if there are any DenseForSparse layers to be updated.

Things to be done:

• get feature_sizes of current data
• persist feature_sizes
• load persisted feature_sizes

[tttthomasssss]

@samsucik assigned as reviewer.