Accelerating Reinforcement Learning of Robotic Manipulations via Feedback from Large Language Models