DINOV3: Self-supervised learning for vision at unprecedented scale