Gaussian Approximation and Multiplier Bootstrap for Stochastic Gradient Descent