Change serialization of embeddings

[rbossy]

JSON has several drawbacks:

• is not standard way to distribute embeddings
• affects performance of the three modules

Recommended:

• native gensim serialization (smaller size, low code volume)

  import gensim
  model = gensim.models.Word2Vec(sentences)
  model.save("path/word2vec.model")
  loadedModel = gensim.models.Word2Vec.load("path/word2vec.model")

• Python shelve (bigger, probably superfast train and predict)

[rbossy]

See 195b98da281da7a1d969fbb6ed867e172fbcd81b

Not tested.

[rbossy]

@mandiayba