Modeling the filtered progress variable source term in lean hydrogen flames using conditional generative adversarial networks