Adaptive Importance Sampling via Stochastic Convex Programming