Training-free P&ID information extraction through a hybrid computer vision and vision language model pipeline for digital twins