Use L2Expanded/L2SqrtExpanded instead of the *Unexpanded variants

[jcrist]

The *Expanded variants were already used in most places, but not everywhere. Using the *Expanded versions should be more performant in all cases, and switching to be uniform across cuml gives more expected behavior.

[dantegd]

It seems that the small numeric changes are making one or a few tests get out of their error margin. We should run some of these tests a number of times (at least 20... more if possible) and see in what margin of error they fall and if it's still an acceptable margin (probably)

[jcrist]

There are 3 consistent failures, all in test_nearest_neighbors_rbc (link).

2 are tolerance issues - bumping atol from 1e-3 to 6e-3 fixes things.

The 3rd is for euclidean with dims=3, n_neighbors=25 - the number of mismatched indices exceeds the threshold (max of 3 is allowed, we now consistently get 37).

If no one is opposed, I think I'll revert the change in pairwise_distances, but keep it in umap/dbscan, and open a follow-up issue to investigate.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Please upload report for BASE (branch-25.04@a2a8080). Learn more about missing BASE report.

Additional details and impacted files

@@               Coverage Diff               @@
##             branch-25.04    #6303   +/-   ##
===============================================
  Coverage                ?   67.07%           
===============================================
  Files                   ?      202           
  Lines                   ?    13076           
  Branches                ?        0           
===============================================
  Hits                    ?     8771           
  Misses                  ?     4305           
  Partials                ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jcrist]

This is all green now, should be good-to-go.