Minimax-Optimal Reward-Agnostic Exploration in Reinforcement Learning