Reduce number of/eliminate xfails in nx-cugraph tests

[rlratzel]

As part of paying down tech debt, there are a number of NetworkX tests that will fail if run using nx-cugraph. These are tests that represent differences between implementations that don't need to prevent users from using nx-cugraph (they're mainly corner cases, such as differences in randomly-selected values for example), but we should still try to reduce these, with the ultimate goal being running the NetworkX tests with nx-cugraph without special cases to dynamically add xfails.

Removing the xfails will involve a combination of updating nx-cugraph and changing the NetworkX tests to be more accommodating of other implementations.

[eriknw]

I'm gathering some notes here.

xfail:

• reason=sssp_path_different: perhaps some networkx tests can be updated to accept any valid path; see nx.all_shortest_paths which computes all valid shortest paths between source and target
• "test_louvain.py:test_undirected_selfloops": we should understand and update how networkx now handles self-loops; see adding a formula that ignores self-loops at the each level of directed louvain algorithm networkx/networkx#6630

We should also look at where we set @networkx_algorithm(is_different=True):

• betweenness_centrality with k can be updated to use the same initial nodes as networkx. This requires updating PLC to acceps a list of nodes instead of a random state seed.

Other technical may be found by git grepping e.g. TODO, FIXME, PERF, DRY (but mostly TODO). Also, functions that define can_run.