Use SequentialReductionKernel for tree-reduction as well

1. Renamed misspelled variable
2. If reduction_nelems is small, used SequentialReductionKernel
   for tree-reductions as it is done for atomic reduction
3. Tweak scaling down logic for moderately-sized number of elements
   to reduce.

   We should also use max_wg if the iter_nelems is very small (one),
   since choosing max_wg for large iter_nelems may lead to under-
   utilization of GPU.