When mitigating bias is unfair: multiplicity and arbitrariness in algorithmic group fairness