Investigate single kernel last block reduce performance

After adding the nondeterministic reduce #4961, which uses a single kernel, we saw that the best performance improvements come from smaller input sizes due to using a single kernel instead of the two kernels used in the run to run determinism reduce.

There is another implementation of reduce that uses a single kernel we should look into, where a single block is designated to reduce the partial aggregates. It will be selected according to an atomic counter. Is this kernel run-to-run deterministic? If so, can it improve performance as well?