Optimal Rates for Averaged Stochastic Gradient Descent under Neural Tangent Kernel Regime