Representation Learning via Consistent Assignment of Views over Random Partitions