Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning