Deep learned process parameterizations provide better representations of turbulent heat fluxes in hydrologic models