Guiding Two-Layer Neural Network Lipschitzness via Gradient Descent Learning Rate Constraints