The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Pre-trained Language Models