SafeAR: Towards Safer Algorithmic Recourse by Risk-Aware Policies