Removing Hidden Confounding in Recommendation: A Unified Multi-Task Learning Approach