Provable Partially Observable Reinforcement Learning with Privileged Information