Masked Feature Prediction for Self-Supervised Visual Pre-Training