GC-VLN: Instruction as Graph Constraints for Training-free Vision-and-Language Navigation