[FEA] try to use cudf::detail::reduce_by_key

[revans2]

Is your feature request related to a problem? Please describe.
There are a few places in spark-rapids-jni where we use thrust::reduce_by_key. As a part of rapidsai/cudf#20860 cudf added cudf::detail::reduce_by_key. We should look into using that instead and see if it speeds things up.

[mythrocks]

:/ On the other hand, CUDF deems cudf::detail as non-public, and ideally not usable from Spark RAPIDS JNI.

I wonder if CUDF might consider a cudf:: version of the same function, or upgrade it to the cudf namespace.