POLICEd RL Learning Closed-Loop Robot Control Policies with Provable Satisfaction of Hard Constraints