Use a custom radix sort implementation instead of qsort in dnxhd encoder.

This is mainly to avoid test failures due to implementation-defined
behaviour of qsort when elements are equal, giving different results
for each of FreeBSD, Linux/glibc and Solaris.
In addition it is about 35 % faster, effect on overall speed is minimal
though (< 2%).
Regression tests are unchanged (i.e. identical to Linux/glibc).

Originally committed as revision 19949 to svn://svn.ffmpeg.org/ffmpeg/trunk