Testing:
o  Giant runs.
o  Records of odd lengths.
o  Empty files.

Improvements:
o  Clean up data types and make sure they cannot overflow. (size_t vs. u64 vs. sh_off_t vs. uns)
o  Switching between direct and normal I/O. Should use normal I/O if the input is small enough.
o  How does the speed of radix splitting decrease with increasing number of hash bits?
   Does it help to use more bits than we need, so that we sort less data in memory?
o  Log messages should show both original and new size of the data. The speed
   should be probably calculated from the former.