Distributionally Robust Offline Reinforcement Learning with Linear Function Approximation