Vision Models Are More Robust And Fair When Pretrained On Uncurated Images Without Supervision