Diverse Randomized Value Functions: A Provably Pessimistic Approach for Offline Reinforcement Learning