Changes for future releases:

* fix unittest for LDA; due to the model's random initiazation, it sometimes 
  returns different result than expected on the testing toy corpus.
* finish tutorials
* implement LDA differently; see Asuncion: "On Smoothing and Inference for Topic Models"
* and the big one: find a reasonable framework for distributed computing in Python
  and translate algorithms into their distributed variants. Avoid Java frameworks
  and over-complex stuff; the main design principles of gensim must still hold.
