Consider advanced libraries for GPFA

[mdenker]

Test if advanced libraries can be used instead of home-brew implementations of certain algorithms (or keep them as fall-back)

[alperyeg]

E.g. check if scikit-learn's FA is exchangeable or compatible with the existing code.
We should make a list here which functions we could exchange in that case.

[Junji110]

fastfa() used in gpfaEngine.py can be replaced by sklearn.decomposition.FactorAnalysis(), which is faster than fastfa().
I've confirmed that the final result of GPFA is essentially not affected by this change.

[alperyeg]

Thanks @Junji110. That sounds very promising, then we should have a version with sklearn and as a fallback a version with the native implementation. For the latter we can split the code in a new module and call it if sklearn is not installed (the usual if HAVE_SKLEARN: do FactorAnalysis else: do fastfa().

[Junji110]

I've also found that minimize() in learnGPparams.py can be replaced by scipy.optimize.minimize(), which is slightly faster than the home-brew minimize().
This replacement has another merit in addition to the speed-up, because the home-brew minimize() was not written by Byron but he adopted a code written by somebody else (Carl Edward Rasmussen, seems to be a professor in Cambridge) under an original licence (see the comment in minimize.py).
Replacing this with the scipy version would prevent us from possible copyright problems.

[Junji110]

I've applied the suggested changes to the codes under gpfa_elephant.