Few-shot Instruction Prompts for Pretrained Language Models to Detect Social Biases