Bugfix

Estimating the intercept only once as the (weighted) average negative gradient, instead of updating it in each boosting step. This prevents predictors that are almost constants from competing with the intercept term.