Provable Interactive Learning with Hindsight Instruction Feedback