Optimize point clouds

[emilk]

In the nyud example, SceneSpatial::load_objects goes from 8.5ms to 16ms.

Classic:

Arrow:

TODO:

• Improve crates/re_query/benches/query_benchmark.rs with benchmarks of:
  • full primary and full secondary
  • full primary and sparse secondary
  • full primary and splat secondary
  • full primary and missing secondary
• Realistic point cloud benchmark, similar to nyud and/or colmap, using actual Points3DPart::load
• Write tests for ComponentJoinedIterator:
  • full primary and full secondary
  • full primary and sparse secondary
  • full primary and splat secondary
  • full primary and missing secondary
• Optimize ComponentJoinedIterator:
  • Special case for the common case when the instance keys are identical
    (no need to compare them - if they are same length, and starts and end on the same key, with the length as the difference, then they are the same)

[emilk]

Some of this optimization work will be easier once we remove the "classic" store

[emilk]

Thanks to #1064 we are now down to 9 ms, which is within the noise of the classic store:

…but I still think we should spend a few hours making a benchmark and see if there are any low-hanging fruit in ComponentJoinedIterator

[emilk]

• Closing in favor of Support huge point clouds in the viewer #1136