x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3

Originally committed as revision 6724 to svn://svn.ffmpeg.org/ffmpeg/trunk