MOFI: Learning Image Representations from Noisy Entity Annotated Images