Towards Understanding the Benefit of Multitask Representation Learning in Decision Process