Assessing the potential of deep neural networks for emulating cloud superparameterization in climate models under real geography boundary conditions