TACO: Temporal Latent Action-Driven Contrastive Loss for Visual Reinforcement Learning