Adaptive Sampling for Stochastic Risk-Averse Learning