Visual Reinforcement Learning with Self-Supervised 3D Representations