Refine fallback kernel for segmented sort

Currently, `cub::DeviceSegmentedSort` has a fallback kernel, that [apply](https://github.com/NVIDIA/cub/blob/0430cc0bfcb7c2496b42da754c215c9b5df8856b/cub/device/dispatch/dispatch_segmented_sort.cuh#L169) different algorithms for different segment sizes. In particular, medium-size segments are sorted by merge sort. If segment doesn't fit into registers, it's sorted by in-shared-memory block radix sort. Otherwise, expensive in-global-memory radix sort is applied.

The suggestion was made, that we have to add small-segment branch and potentially experiment with refining the granularity at which we assign threads to segments at fallback kernel.

Read cub/device/dispatch/dispatch_segmented_sort.cuh around the DeviceSegmentedSort fallback kernel at line 169, then trace how segment sizes select the existing algorithms. Determine how to add the suggested small-segment branch and evaluate finer thread-to-segment granularity; done means the fallback strategy has been refined based on those experiments.

Written by the indexing model from the issue text.