nx-cugraph: add SSSP (unweighted)

[eriknw]

There are many more traversal algorithms to implement, but these get us started!

[eriknw]

The errors are interesting:

RuntimeError: non-success value returned from cugraph_bfs: CUGRAPH_UNKNOWN_ERROR std::bad_alloc: out_of_memory: CUDA error at:

We are calling plc.bfs repeatedly as part of all_pairs_shortest_path_length, but shouldn't this work still? Shouldn't memory be cleaned up? All results are turned into Python dicts and they shouldn't be on device anymore.

Heh, so it looks like single_source_shortest_path_length is a good example of a networkx algorithm that other networkx algorithms call in a loop. It's used by (at least) these:

• all_pairs_shortest_path_length
• closeness_centrality
• bipartite.closeness_centrality
• incremental_closeness_centrality
• average_shortest_path_length

[rlratzel]

/merge