ARROW-14052: [C++] Add approximate_median aggregation

[lidavidm]

This wraps the tdigest aggregation to make it easier for bindings to use. This PR is on top of ARROW-14050 (#11199).

[github-actions]

https://issues.apache.org/jira/browse/ARROW-14052

[ianmcook]

@lidavidm When I test the hash aggregate variation of this this from the R bindings on arrays of integer data that have integral medians within groups, I'm seeing errors like this:

Error: Invalid: Referenced field x was double but should have been int32

Do these kernels always return float64 output or are there some cases where integer input will give integer output?

[lidavidm]

@ianmcook hmm, no, it should always give float64 regardless of the input type - I'll take a look tomorrow (if you push it somewhere I can dig into it too)

[ianmcook]

@ianmcook hmm, no, it should always give float64 regardless of the input type - I'll take a look tomorrow (if you push it somewhere I can dig into it too)

I built Arrow C++ from this PR and I'm using that to test the changes to the R bindings in #11018. The test that fails with this error is here: https://github.com/apache/arrow/pull/11018/files#diff-a18898eb61bfc3d3dda292cf594b26d5f08a8cd7b455b72ed9c58f6e9c1311edR234-R249

[lidavidm]

Thanks, that should be fixed now - I had declared the wrong output type, leading to expressions getting bound with the wrong types. The tests pass locally for me now.

[ianmcook]

Thanks, that should be fixed now - I had declared the wrong output type, leading to expressions getting bound with the wrong types. The tests pass locally for me now.

Yep that fixes it, thanks!

[ianmcook]

Is this ready to merge? The CI failure appears to be unrelated.

[lidavidm]

I think we should be good here? We can tackle anything in a followup if needed.