The Role of Inherent Bellman Error in Offline Reinforcement Learning with Linear Function Approximation