[BUG] Check if Dask has quit to avoid throwing an exception and trigg…

…ering a segfault on ddp exit (#3961)

Currently, when training with ddp, if dask exits before the `CuGraphStore` is cleaned up, an exception is thrown, which causes ddp to quit with an error, which then causes a segfault, making users think that the workflow has failed when it has actually succeeded.  This bug gracefully displays a warning if the dask dataset can't be deleted, which resolves this issue.

Authors:
  - Alex Barghi (https://github.com/alexbarghi-nv)

Approvers:
  - Vibhu Jawa (https://github.com/VibhuJawa)
  - Tingyu Wang (https://github.com/tingyu66)
  - Rick Ratzel (https://github.com/rlratzel)

URL: #3961

python/cugraph-pyg/cugraph_pyg/data/cugraph_store.py

@@ -320,7 +320,13 @@ def __init__(
     def __del__(self):
         if self.__is_graph_owner:
             if isinstance(self.__graph._plc_graph, dict):
-                distributed.get_client().unpublish_dataset("cugraph_graph")
+                try:
+                    distributed.get_client().unpublish_dataset("cugraph_graph")
+                except TypeError:
+                    warnings.warn(
+                        "Could not unpublish graph dataset, most likely because"
+                        " dask has already shut down."
+                    )
             del self.__graph
 
     def __make_offsets(self, input_dict):