The Dimension Strikes Back with Gradients: Generalization of Gradient Methods in Stochastic Convex Optimization