Improved Scalable Lipschitz Bounds for Deep Neural Networks