Improve GroupOrderingPartial performance #14565

The current implementation of GroupOrderingPartial updates its internal state by converting any incoming batch into the row format, and then traversing that format to determine if the sort key changed.

I believe that the same can be computed (more effiently?) without such a conversion by using arrow_ord::partition::partition.