Policy Iteration with Human Feedback: Bringing Post-Training RL to In-context Learning