Implement Gradient Boosted Feature Selection

[tomtung]

as described in http://alicezheng.org/papers/gbfs.pdf

Basically it seems like: when measuring the goodness of a tree node split (for learning a regression tree), add a tunable constant penalty if the feature used for splitting hasn't been used by existing trees.

Seems quite simple, and can be a nice alternative for people who use feature importance to select features.

[tqchen]

Let us move roadmap issues to #574 so it can be tracked in a more easier way