The False Promise of Zero-Shot Super-Resolution in Machine-Learned Operators