Feature selection

[chbrown]

Provide feature selection results within the current array of feature functions, based on our labeled data.

We particularly want to avoid disaster-type bias; a multitude of features along with the sparse data we're training on means over-fitting is an big concern (intuitively, at least— cross-disaster evaluation is in the works).