Learning Subgrid-scale Models with Neural Ordinary Differential Equations