A stalled #85519 adds capabilities to dump cuda graphs for inspection. We would like to extend it to

1. provide a way to get all nodes from the graph
2. modify nodes to patch in addresses for the inputs to avoid copying inputs. Note, this is more complicated than just comparing old and new pointers and swapping one for another because kernels could operate on input slices, so one would need to inspect if kernel argument falls within the range of the original input tensor and, if so, replace it with the corresponding new address + offset.

cc @mcarilli @ezyang @eqy