High Probability Convergence of Stochastic Gradient Methods