Option of numerical and not just one-hot features

[rkurchin]

I'd like to play with this. My suspicion is that if it works, it will take a lot more epochs to train to the same accuracy, but it would also allow us to get away with much smaller models. In addition, having input parameters that are actual values would allow cool things like very transparent sensitivity analyses via autodiff.

This would almost certainly require some normalization of the input features; so the AtomFeat objects would need to store the normalization in order to be able to invert the encoding properly. My inclination now is that this is best achieved via a new type (e.g. split into OneHotAtomFeat and NumericalAtomFeat that both inherit from an abstract AtomFeat class? A lot of things could be fairly easily dispatched onto both)

[rkurchin]

This is resolved in a basic way by DirectCodec added in #118