cub::DeviceRadixSor works with a hack:

- Code compiles and a test in tests/cuv_sort.cu outputs a correct answer

HACK: combined Topkey, key into a single unit64_t.
- This is to avoid using a struct with two uint32_t members. Which is not supported by cub::DeviceRadixSort::SortPairs.
- TODO: use `thrust::sort_by_key` instead of `cub::DeviceRadixSort::SortPairs`

Author: qezlou