The Thinking Leakage in Reinforcement Learning for LLMs