CC: @ndawe

I have been experimenting with our latest implementation of AdaBoost and it is only now that I actually become aware that it is quite slow to build trees.

The problem is that it recomputes X_argsorted for every tree, which is a complete waste of computing power since that doesn't change during the construction process. It could be computed once and for all, just like we do in random forests.

This is the price to pay for having a generic implementation of AdaBoost that can have as input any kind of base estimators. However, since AdaBoost is likely to be used with trees, I'd be in favor of adding specialized optimization routines for that case.

What do you guys think?