Attention Speaks Volumes: Localizing and Mitigating Bias in Language Models