Policy-Induced Self-Supervision Improves Representation Finetuning in Visual RL