Data-driven calibration of linear estimators with minimal penalties