One Form to Transfer Them All: Pretraining Multilingual Language Models Beyond Native Orthography