Benefits of Learning Rate Annealing for Tuning-Robustness in Stochastic Optimization