An All-MLP Architecture for vision