Overhaul MultivariateExpansion #338
Description
This is now an old conversation, but in 1.0 we (matt) figured out how to create maps that were efficient to evaluate and invert. In 2.0 I mostly focused on creating the ability to train and research uses of maps. In the future (e.g. 3.0) I want to get it to a point where we are able to research and experiment with different parameterizations of the maps themselves (this was elegantly put by Matt in our last conversation).
This entails making MultivariateExpansion and the like more friendly to tweak and call from the binding language and, additionally, creating the ability to call the binding language from C++ for univariate or multivariate members of the expansion. Then, we don't have to worry as much about implementing things in straight C++ efficiently the first time around, and we can really experiment with different abilities. As an example, it would be really cool/nice to be able to just throw in a Pytorch/Tensorflow neural network in a component and be able to use it for our map.