cVLA: Towards Efficient Camera-Space VLAs