Trains GBM and LightGBM iteratively, numpy fixes

LightGBM and sklearn's gbm now use warm_starting or iterative training to find the best number of trees