Contrastive Error Attribution for Finetuned Language Models