From ImageNet to Image Classification: Contextualizing Progress on Benchmarks