Tighter Generalization Bounds for Iterative Differentially Private Learning Algorithms