Optimizing Neural Networks with Gradient Lexicase Selection