[ENH] Add MG neighborhood sampling to cugraph Python API

[rlratzel]

• cugraph API signature discussed and defined
• cugraph tests in place
• cugraph docstring
• cugraph API implementation defined (calling pylibcugraph API stub)
• pylibcugraph tests in place
• pylibcugraph docstring
• pylibcugraph API implementation defined (calling C API stub)
• replace stubs with actual calls everywhere when C API finished, ensure tests with actual data pass

This depends on: #2073

---

Notes

• A separate graph service for use by DGL would operate in a more parallel batch-type mode by
  • Collecting several calls from different clients and assigning each client a unique label (an integer type) and associating the starting vertices for that client with that label
  • The starting vertices arrays are combined into a single flat array of tuples of (starting_vertex, label)
  • When enough client calls are received to exceed some threshold, the service makes a single call to the pylibcugraph/C API uniform_neighborhood_sample() with the array of tuples.
  • The uniform_neighborhood_sample() is run (multi-GPU) and the result is processed. The result is a flat list of tuples of tuples: ((src, dst, label, index), rx_counts) where index is used to reconstruct the path (eg. if index=n, that edge is then the nth edge in the path for that label) and rx_counts is ???
• Let start with this API for cugraph:

cugraph.dask.sampling.uniform_neighborhood(G, start_info_list, fanout_vals, with_replacement=True)

where G is a cugraph Graph instance created using from_dask_cudf_edgelist()

• Lets start with this API for pylibcugraph:

pylibcugraph.uniform_neighborhood_sampling(G, start_info_list, fanout_vals, with_replacement=True)

where G is a MGGraph instance.