Incorporating Bias-aware Margins into Contrastive Loss for Collaborative Filtering