On the Convergence of Stochastic Multi-Objective Gradient Manipulation and Beyond Shiji Zhou