Hypervolume Maximization: A Geometric View of Pareto Set Learning