Guiding Large Language Models via Directional Stimulus Prompting