[BUG] SCC results for graph with loops seems incorrect and does not match Nx nor SciPy

[rlratzel]

I recently observed behavior from cugraph.strongly_connected_components when handling directed graphs that have loops that appears to be a bug. The same input given to the SCC algos in SciPy and NetworkX both return identical results which match expectations (the input graph is trivially small and easy to predict results with, and is included in the test script attached below).

Other information:

• This was initially discovered through using our C++ API from libcugraph directly, so that should eliminate the root cause of this issue as something in the python layer.
• The handling of loops are assumed to be related, but only because the attached script uses them in the comparison and that may not be the root cause.
• The attachments describe the issue best, but to summarize:
  For this input:

[[1., 1., 0., 0., 0.,],
 [1., 1., 0., 0., 0.,],
 [1., 0., 1., 0., 0.,],
 [0., 1., 0., 1., 0.,],
 [0., 0., 0., 0., 1.,],
]

this output is observed:

NUM LABELS: sp: 4, cu: 2, nx: 4
GROUPS: sp: [{0, 1}, {2}, {3}, {4}], cu: [{0, 1, 2, 3}, {4}], nx: [{0, 1}, {2}, {3}, {4}]

where sp is SciPy, cu is cuGraph, and nx is NetworkX. Each library's strongly connected components API was called using a directed graph created from the array above.

connected_components_example_py.txt

connected_components_example_output.txt

[rlratzel]

FYI: @anaruse