Learning Smooth Humanoid Locomotion through Lipschitz-Constrained Policies