add dynamic topic modeling to clustering models

[rbroc]

fixes #11

Need to:

• add fit_transform_dynamic to ClusteringTopicModel
• add tests for dynamic models

Still WIP and not tested, need to discuss what to do to compute temporal components when feature importance is based on distance from topic centroid, i.e.:

• are topic centroids computed based on all documents, or only documents present in a given time bin?
• what to do with terms that do not occur in the documents? set similarity to 0 or -1 for terms that do not occur within a given time bin, or use all terms?

[rbroc]

This might be ready for a first review.

Couple of notes when feature_importance='centroid':

• We decided to compute topic centroids based only on documents present in a time slice. When no documents are present, the resulting vectors is all nans. For the purpose of computing feature importances, vectors are set to zero in a slightly hacky way because sklearn.metrics.pairwise_distances does not take non-finite values with predefined metrics;
• In the topic-term matrix, I set values for both topics for which no documents are available and terms that are not present in that time slice to NaN. We might want to discuss whether this is what we want, or whether zeros are better (perhaps for consistency with other feature_importance options)

Docs deployment is failing, but this is a separate issue #30.

[rbroc]

oh also there's a few redundant commits because I accidentally pushed some unrelated stuff, let me know if you'd rather have a cleaner PR