Port search_documents_by_keywords to C-Top2Vec

[CodingKoopa]

This PR enables support for searching documents by keyword for C-Top2Vec models.

The first commit adds a Top2Vec parameter named combine_ngram_vocab to add both single words and n-grams to the phrase embeddings. If this is enabled, then the original behavior described in #364 is restored. The motivation is we don't use the top topic words (and if we do, we are manually selecting representative words/phrases), but we do want to be able to search by both word or phrase.

The second commit enables search_documents_by_keywords() to work with C-Top2Vec. This relies on the same self.document_vectors member variable as original Top2Vec, which here is the multi-vector document representation. Since the splitting of the original documents is a byproduct of the sliding window average, I have introduced another member self.multi_document_labels to maintain the mapping of each multi-document vector to the original vector.