From Multitask Gradient Descent to Gradient-Free Evolutionary Multitasking: A Proof of Faster Convergence