[WIP] Sparse semiring cleanup + hash table and batching strategies

[cjnolet]

This branch includes several new features and optimizations:

1. Introduces a hash table strategy to sparsify the vector in the coo spmv shared memory
2. Adds cucollections to build (necessary for item 1)
3. Adds a batching strategy for rows with nnz too large to fit into shared memory
4. Removes the need for the cusparse csrgemm
5. Uses raft handle in distances_config_t rather than accepting each resource explicitly
6. Removes the naive CSR semiring code

This PR is also required to merge #261, which introduces the remaining distances

[divyegala]

Initial review looks good to me (excepting strategies part which I will clean up soon)! Only one comment and one question - the tests for the new distances remain to be added, right?

[cjnolet]

@dantegd, I've updated this PR for branch 21.06 but I'm not familiar with the new package management. I'm getting the following build error, though I do see cuco in the build directory. Is there something we need to add to the gtest build?

In file included from /home/cjnolet/workspace/raft/cpp/include/raft/sparse/distance/coo_spmv.cuh:20,
                 from /home/cjnolet/workspace/raft/cpp/test/sparse/dist_coo_spmv.cu:28:
/home/cjnolet/workspace/raft/cpp/include/raft/sparse/distance/coo_spmv_strategies/hash_strategy.cuh:21:10: fatal error: cuco/static_map.cuh: No such file or directory
   21 | #include <cuco/static_map.cuh>
      |          ^~~~~~~~~~~~~~~~~~~~~

[cjnolet]

Just marking this here so I don't forget- It looks like the current cuco API is slightly different than it was when the hash_table in this PR was written. I see two differences so far that are causing compile errors:

1. The slot_type was removed from the static_map::device_mutable_view and it's unclear to me which one should be used in its place. I tried using from pair_atomic_type from static_map, which seems to remove this compile error but I'm still not sure if it's the correct one to use
2. make_from_uninitialized_slots seems to have been removed from static_map altogether and it's not clear to me what should be used in its place.
3. A seed argument was added to a constructor and I'm assuming this might need to be incorporated into cucollections.

These are the obvious errors I see at the top of the trace. I'm hoping fixing these might fix everything else? I'm not sure yet, but I think maybe we should push this change to 21.08.

[cjnolet]

Closing this as it's replaced by #269