Adds fail_on_nonconvergence option to pagerank to provide pagerank results even on non-convergence

[rlratzel]

closes #3613

Prior to this PR, pagerank will raise a RuntimeError if it fails to converge, often because the max_iter param is set too small (intentionally or otherwise). This PR adds the optional paramter fail_on_nonconvergence which defaults to True (ie. the current behavior to ensure backwards-compatibility) that allows a caller to run pagerank and get results even if it did not converge. When fail_on_nonconvergence is False, pagerank will return a tuple containing the pagerank results and a bool indicating if the results converged or not).

[VibhuJawa]

Approving from the Python/Dask cugraph layer.

[seunghwak]

LGTM except for one additional complaint.

[seunghwak]

Shouldn't this better be std::optional<device_span<>>?

[ChuckHastings]

Added #3659 to address this in a separate PR.

[ChuckHastings]

Nevermind. Addressed in this PR since we had to make other changes.

[eriknw]

For user-facing API, I wonder whether fail_on_nonconvergence is the clearest and most convenient:

pagerank(..., max_iter=3, fail_on_nonconvergence=False)

I think I would prefer a more direct, affirmative argument, such as:

pagerank(..., num_iter=3)

[eriknw]

An alternative implementation to this could be something like:

import operator as op

...

result_tuples = client.map(convert_to_return_tuple, cp_arrays)
meta = cudf.DataFrame(columns=["vertex", "pagerank"])
meta = meta.astype({"pagerank": "float64", "vertex": vertex_dtype})
ddf = dask_cudf.from_delayed(client.map(op.itemgetter(0), result_tuples), meta=meta).persist()
converged = client.submit(all, client.map(op.itemgetter(1), result_tuples)).result()

[VibhuJawa]

Oh Nice, Did not know we could do op.itemgetter like this. Very cool to learn. Thanks

[ChuckHastings]

/merge