Exploiting random projections and sparsity with random forests and gradient boosting methods -- Application to multi-label and multi-output learning, random forest model compression and leveraging input sparsity