Generalizing Fairness using Multi-Task Learning without Demographic Information