Label Propagation Training Schemes for Physics-Informed Neural Networks and Gaussian Processes