Stochastic Conditional Gradient Methods: From Convex Minimization to Submodular Maximization