Improvements in matrix::gather: test coverage, compilation errors, performance

[Nyrio]

In order to deprecate copy_selected from ann_utils.cuh, I wanted to make sure that the performance of matrix::gather was on par. But in the process I discovered that:

• Map transforms and conditional copy were not tested at all.
• In fact, most of the API in gather.cuh wasn't covered in tests and some of the functions didn't even compile.
• The same type MatrixIteratorT was used for the input and output iterators, which made it impossible to take advantage of custom iterators, as is needed in kmeans_balanced to convert the dataset from T to float and gather in a single step.
• The performance was really poor when D is small because the kernel assigns one block per row (so a block could be working on only 2 or 3 elements...)

This PR addresses all the aforementioned issues.

[Nyrio]

Below is a comparison of before/after performance. The speedups range from 1x to 7x in my benchmark (possibly more with an even smaller out_cols).

[codecov-commenter]

Codecov Report

Base: 87.99% // Head: 87.99% // No change to project coverage 👍

Coverage data is based on head (029724f) compared to base (187ff9e).
Patch has no changes to coverable lines.

Additional details and impacted files

@@              Coverage Diff              @@
##           branch-23.02    #1126   +/-   ##
=============================================
  Coverage         87.99%   87.99%           
=============================================
  Files                21       21           
  Lines               483      483           
=============================================
  Hits                425      425           
  Misses               58       58           

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[tfeher]

Hi Louis, thanks for the PR, it looks good! While we are here, could you improve a little bit the docstring of gather, I have left a few questions.

[Nyrio]

I'm not sure what's up with CI but it doesn't seem to be a compiler nor test error.

[tfeher]

Thanks Louis for updating the doc. The PR looks good to me.

[cjnolet]

@Nyrio the changes look good here. Before we merge this, can you verify that the changes to the template params don't break cuml downstream?

[Nyrio]

@cjnolet There is one call to gather_if in cuML and it shouldn't break (actually, the only thing that can potentially break with my change is the mdspan overload if all the template args are explicitly specified).

[cjnolet]

LGTM

[cjnolet]

/merge