Physics-informed deep neural networks applied to scalar subgrid flux modeling in a mixed DNS/LES framework