Surfacing Biases in Large Language Models using Contrastive Input Decoding