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