Add cuco::static_multimap

[PointKernel]

Closes #59

Added cuco::static_multimap:

• Major performance optimizations:

  • Relaxed atomic and reinterpret_cast (to be replaced by atomic_ref once it's ready)
  • Shared memory buffer at CG/warp/block level
  • Double hashing instead of linear probing
  • Use of CG with parallel retrieval
  • Vector loads (each thread loads two consecutive slots) for 32-bit integers
    

• Comprehensive benchmark to evaluate multimap performance

  • nvbench instead of google benchmark
  • Jupyter notebooks showing benchmarking results (cuCollections/benchmarks/analysis/notebooks)

• Flexible switch between vector/scalar loads and between different probing methods

  • class ProbeSequence as a template parameter for cuco::static_multimap
  • By default, using per-block shared memory buffer + vector loads for 32-bit integers while per-CG buffer + scalar loads for 64-bit integers

[abellina]

Just checking in, what is the status of this PR?

[jrhemstad]

Just checking in, what is the status of this PR?

Waiting on some final review changes.