Bugfixes

• Estimating the intercept only once as the (weighted) average training response, instead of updating it in each boosting step. This prevents predictors that are almost constants from competing with the intercept term.
• Fixed a minor bug related to min_observations_in_split.