MUA-RL: Multi-turn User-interacting Agent Reinforcement Learning for agentic tool use