On the upper bound of the prediction accuracy of residue contacts in proteins with correlated mutations: the case study of the similarity matrices