What/Where to Look Next? Modeling Top-Down Visual Attention in Complex Interactive Environments