Approximating Interactive Human Evaluation with Self-Play for Open-Domain Dialog Systems