Learning without the Phase: Regularized PhaseMax Achieves Optimal Sample Complexity