replace MULL by MULH + shift in dct32() (code is very slightly faster, and the compiler should be able to optimize the shift away on 64bit archs)

Originally committed as revision 6045 to svn://svn.ffmpeg.org/ffmpeg/trunk