Descending through a Crowded Valley - Benchmarking Deep Learning Optimizers