Benefits of Early Stopping in Gradient Descent for Overparameterized Logistic Regression