Removing trees from a model

[ldesreumaux]

As a follow up to #2281, I think that it would be great if XGBoost supported the removal of trees from a model.

It would allow the implementation (with callbacks) of adaptive learning rate strategies that need to abandon some trees before adjusting the learning rate. More generally, it would allow to implement various "ensemble pruning" methods.

[trivialfis]

It's doable with process_type as update, but it's quite hacky right now.

[trivialfis]

I have a local branch that can slice the model, but it still need to handle Python's negative slice.