Scaling and Benchmarking Self-Supervised Visual Representation Learning