Towards Data-Free Model Stealing in a Hard Label Setting