Recovering Petaflops in Contrastive Semi-Supervised Learning of Visual Representations