User Feedback in Human-LLM Dialogues: A Lens to Understand Users But Noisy as a Learning Signal