Replace ner_features by text_dense_features

[tabergma]

If you wanted to use additional features in the CRFEntityExtractor you had to add ner_features to the feature list of the entity extractor. Additionally, you need a featurizer that actually sets those features. The main difference of those features are, that they should come as a sequence.

With the upcoming changes it will be possible to easily define for every featruizer whether it should return its features as sequence or not. Also, we are planning to add more featurizers that return embeddings from pre-trained language models.

To be able to easily make use of any of those features, we should exchange the logic of ner_features. You should be able to use the features any featurizer, that returns word embeddings, in the CRFEntityExtractor. Those features will be stored in the attribute text_dense_features of the message. The featurizer needs to return a sequence of features. Otherwise, the features cannot be used and a warning should be printed.