Emergent Active Perception and Dexterity of Simulated Humanoids from Visual Reinforcement Learning