Neuroevolution Surpasses Stochastic Gradient Descent for Physics-Informed Neural Networks