Demystifying SGD with Doubly Stochastic Gradients