Fix infinite diff in LdaModel.do_mstep

[horpto]

Fix partially #416
Fix partially #2051

This PR solves infinite diff in output. It works for me. This caused by a big negative values in self.state.sstate and narrow dtype of values. I try to solve only this bug. self.expElogbeta can contain zeros anyway. That can be solved by pointing a more wide dtype, for example np.float64. I don't increase dtype of self.expElogbeta as I know there is a drawback between memory consumption/file sizes/sent bytes and precision. User should be smart enough and choose wisely himself.

[menshikh-iv]

ping @johann-petrak @stevemarin @TC-Rudel

Guys, this change should fix diff=inf/nan issue (from #416), can anyone check is current PR fixed your problems:

1. clone horpto fork: git clone git@github.com:horpto/gensim.git && cd gensim
2. install gensim from inf-mstep branch: git checkout inf-mstep && pip install -e . && python setup.py build_ext --inplace
3. check that diff=nan doesn't appear for you any more?

[menshikh-iv]

Thanks @horpto