How Infinitely Wide Neural Networks Benefit from Multi-task Learning -- an Exact Macroscopic Characterization