Bagging in overparameterized learning: Risk characterization and risk monotonization