Two halves make a whole: A novel grid–graph contrastive alignment framework of spatiotemporal features in traffic flow prediction