Scalable Back-Propagation-Free Training of Optical Physics-Informed Neural Networks