Text this: Discrete Physics-Informed Training for Projection-Based Reduced-Order Models with Neural Networks