[ENH] Implement HITS using the graph primitives

[ChuckHastings]

Our current implementation of HITS uses the gunrock implementation which is limited to 1 CPU.

Inherently HITS is the same computational pattern as Pagerank and Katz Centrality. It should be possible to implement HITS using the graph primitives we use for Pagerank and Katz, giving us a HITS computation that can scale to multiple GPUs.

Suggest a sequence of PRs for this:

• Move the existing C++ implementation into a cugraph::legacy namespace, updating the python to reference the legacy implementation. Define the new C++ API, define C++ unit tests using the C++ API.
• Implement the algorithm in C++, all unit tests should pass
• Implement the cython/python changes to support the new API
• Delete the legacy implementation

[seunghwak]

To give little more detail about the implementation strategy,

Compare

https://github.com/networkx/networkx/blob/main/networkx/algorithms/centrality/katz.py#L172

        for n in x:
            for nbr in G[n]:
                x[nbr] += xlast[n] * G[n][nbr].get(weight, 1)

https://github.com/networkx/networkx/blob/main/networkx/algorithms/link_analysis/hits_alg.py#L116

        for n in h:
            for nbr in G[n]:
                a[nbr] += hlast[n] * G[n][nbr].get("weight", 1)
        # now multiply h=Ga        
        for n in h:            
            for nbr in G[n]:
                h[n] += a[nbr] * G[n][nbr].get("weight", 1)

The first part of the HITS's core implementation is practically identical to katz, but HITS has the second part as well.

Katz implements the first part using
copy_v_transform_reduce_in_nbr (assuming that the matrix is stored transposed)
https://github.com/rapidsai/cugraph/blob/branch-21.08/cpp/src/experimental/katz_centrality.cu#L106

but HITS will require calling both

copy_v_transform_reduce_in_nbr and copy_v_transform_reduce_out_nbr (the order will vary based on whether we assume the graph adjacency matrix is stored as is (row-major) or transposed (column-major). For PageRank & Katz, storing the transposed matrix is beneficial in performance to avoid atomics, but for HITS, there is no benefit in performance as we are using both copy_v_transform_reduce_in_nbr and copy_v_transform_reduce_out_nbr.