A DeepONet multi-fidelity approach for residual learning in reduced order modeling