A priori analysis on deep learning of subgrid-scale parameterizations for Kraichnan turbulence