Sampling from Log-Concave Distributions with Infinity-Distance Guarantees and Applications to Differentially Private Optimization