In-N-Out: Pre-Training and Self-Training using Auxiliary Information for Out-of-Distribution Robustness