Stochastic Polyak Step-size for SGD: An Adaptive Learning Rate for Fast Convergence