Improve aggregation device functors

[PointKernel]

Description

While working on #16619, I noticed that aggregate_row is always instantiated with the same template values, making the template parameters unnecessary. This PR simplifies the function by removing the template parameters and moving the device aggregators to their own header. This is a preparatory step for #16619, where additional overloads of the device aggregators will be introduced.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[mhaseeb123]

Minor comments. LGTM otherwise!

[davidwendt]

Looks good.
I see there are a bunch of cudf::detail::atomic* calls here. Do we have an issue to remove these in favor of cuda::std::atomic?

[PointKernel]

Do we have an issue to remove these in favor of cuda::std::atomic?

Good question.

We have #13575 tracking potential cleanups with cuda::atomic_ref, but it turns out to be unfeasible. atomic_ref is meant for 4B/8B integers and floats, whereas cudf needs to support types like durations, booleans, and timestamps, etc. Therefore, we can't fully get rid of cudf::detail::atomic. I’ll add a comment in #13575 to clarify the scope for future work.

[PointKernel]

/merge