abstract out feature generation

* datasets -> simple features: many to many
* simple features -> complex features: many to many
* features -> ABTs : many to many
* ABTs -> modeling tasks : one to one
* modeling tasks -> models : one to many

It would be nice if I could just tie a folder of features to a dataset the way I do for models, but this would
probably get really messy quickly. If I have 3 raw datasets and I want to calculate log transforms for every column in all three, I should need to make three scrips that do the exact same thing. More likely, I will only want to log transform certain columns, so then maybe it makes sense to have a separate script for each (to specify the columns).

Maybe having a "simple features" folder for each dataset wouldn't actually be so bad. The alternative is to require that a user incorporate features to the pipeline themselves, which we want to minimize as much as possible since it limits our ability to enforce generality.