Replace the MoGe-v1 depth model with MoGe-3 (float32)

Keep weights only: runtime lives in mlx-vlm; refresh card and validation

Add frozen BF16 MLX SAM 3D Objects inference with async streaming

initial commit