[Feature request] Arbitrary base learner

Its pretty cool that I can define my own loss function and gradient for xgboost, and then use the linear, tree, or dart base learners to optimize my loss function.

It'd be really cool if I could specify my own base learner, perhaps in the form of an sklearn class with a fit method, a predict method, and support for sample weights.

It'd really open up a whole new world of possibilities to be able to use the Xgboost algorithm to fit a wider range of possible base learners.