Understanding Stereotypes in Language Models: Towards Robust Measurement and Zero-Shot Debiasing