Zeroth-Order Gradient and Quasi-Newton Methods for Nonsmooth Nonconvex Stochastic Optimization