Improve tests of pLSA and PCA

[ukoethe]

Compare with Michael's matlab implementation, using the same random numbers.

And while you are at it: variable names and Python documentation could also be improved.

[ukoethe]

This may also be an opportunity to add the nearest-neighbor preserving pLSA variant (as well as an analogous non-negative factorization for the Frobenius norm) described in [Cai et al. 11], PAMI 33(8):1548-1560, and perhaps the factorization minimizing the earth mover distance by [Sandler & Lindenbaum 11], PAMI 33(8):1590-1602.

ICA as a further alternative is probably best supported by wrapping one of the existing libraries.