Provable and Practical Online Learning Rate Adaptation with Hypergradient Descent