Theory-guided physics-informed neural networks for boundary layer problems with singular perturbation