RLingua: Improving Reinforcement Learning Sample Efficiency in Robotic Manipulations With Large Language Models