Near-optimal and Efficient First-Order Algorithm for Multi-Task Learning with Shared Linear Representation