Dual Natural Gradient Descent for Scalable Training of Physics-Informed Neural Networks