How to Train Your Long-Context Visual Document Model