Investigating the Role of Transfer Learning in Enhancing CNN-Based Subgrid-Scale Models for Geophysical Turbulence