DrM: Mastering Visual Reinforcement Learning through Dormant Ratio Minimization