The Computational Advantage of Depth: Learning High-Dimensional Hierarchical Functions with Gradient Descent