Second-Order Information in Non-Convex Stochastic Optimization: Power and Limitations