Principal-Agent Multitasking: the Uniformity of Optimal Contracts and its Efficient Learning via Instrumental Regression