Debiasing Text Safety Classifiers through a Fairness-Aware Ensemble