Measuring consistency via ensemble margin and local prediction variability: Auditing decision systems in the presence of predictive multiplicity