Scaling, Simplification, and Adaptation: Lessons from Pretraining on Machine-Translated Text