Merging and Disentangling Views in Visual Reinforcement Learning for Robotic Manipulation