Limiting accuracy of segregated solution methods for nonsymmetric saddle point problems