Dual Cone Gradient Descent for Training Physics-Informed Neural Networks