Generalization of ERM in Stochastic Convex Optimization: The Dimension Strikes Back