Improve RBC eps-neighborhood query performance

[mfoerste4]

This PR significantly improves performance of epsilon neighborhood search via RBC.

Scope:

• the rbc index was modified to contain a reordered dataset allowing for better access patterns
• all kernels for dense, sparse and hybrid have been improved

Optimizations include:

• improve pruning of complete landmarks by pre-fetching distances and minimizing overhead for skipping
• specialized 2D and 3D kernels, allowing for register per-fetch of query points
• improve inner loop iterating landmark neighborhood
  • pruning of points within selected landmark neighborhood using triangle inequality
  • reverse iterate landmark neighborhood to allow complete processing stop once subsequent points cannot be reached
  • removal of shared memory atomics in favor of warp voting
• general prevention of branches within a warp (at least to some extent)

CC @cjnolet , @tfeher

[mfoerste4]

Thanks @Nyrio for the review. i have added a couple of minor updates that slightly improve perf further. Still could not find a way to remove the latency issue of the distance-compute-loads though.

I have not applied the changes to all kernel variants yet until we decide to settle on it.

[tfeher]

Thanks Malte for the PR, it is great to have a large speedup for the eps-neighborhood kernels. Overall the PR looks good to me.

Please add a few more details about the optimizations into the description, eg:

• pruning of points within selected landmark neighborhood using triangle inequality
• specialized 2D and 3D kernels

Many thanks to @Nyrio for helping with the optimizations and the review!

[tfeher]

/merge