DOC: document space/time complexities, where known

[tylerjereddy]

One thing I've noticed while doing hyperparameter search experiments internally with GFDLClassifier is that the space complexity of the Moore-Penrose exact solve with larger neural networks is problematic (i.e., 512 GB RAM is not sufficient to fit to a relatively small dataset). The ridge regression route seems to be a bit more tractable in this regard.

I suspect this all matches with theoretical expectations, but it would be helpful to encode the expected time and space complexities of the algorithms for which this is known in our documentation.

In the short term, this probably means documenting this for Moore-Penrose (i.e., complexity relationships to network size and design matrix). Longer-term, we may also want to add some docs about complexities of some of the Ridge solver options if they are known. This should help the reader/user make informed decisions.

Usually this information would go in the Notes section of a docstring.

[tylerjereddy]

Note that even the Ridge route can be problematic--in hyperparameter optimization experiments with larger design matries and large neural networks the automatically-chosen solver sometimes transits via an SVD that causes a MemoryError.