generalization to support new architectures

[mheilman]

As pointed out by @beckermr, we could probably make a few changes here and there to make this code more useful for implementing a wider range of architectures.

I think we probably want to avoid increasing the scope too much, but some of the functionality for getting TF to work with sklearn isn't specific to classification/regression or MLPs (e.g., pickling, handling dense/sparse input matrices, partial_fit support), and I think a little work on this (e.g., adding some mixins or refactoring the base class) could be worthwhile.

Related to #4.