Boosting the Confidence of Generalization for $L_2$-Stable Randomized Learning Algorithms