Improving Lipschitz-Constrained Neural Networks by Learning Activation Functions