Fighting Contextual Bandits with Stochastic Smoothing