Learning Agent-Aware Affordances for Closed-Loop Interaction with Articulated Objects