PromptEcho: Annotation-Free Reward from Vision-Language Models for Text-to-Image Reinforcement Learning