Trajectory Improvement and Reward Learning from Comparative Language Feedback