LDA. Seed topic-words matrix

[IvanAntipov]

The widely used practice of LDA is the seeding of the matrix of topic-word

ex.
https://github.com/vi3k6i5/GuidedLDA

This possibility required to:

1. Have stable clustering (topics number remain if data is not changed significantly) when re-training the model
2. Have semi-supervised model (when user define some words for topics)

This is a very common business requirement