Quantifying and mitigating the impact of label errors on model disparity metrics