more accurate mmx dct (dont discard bits for fun)
fixing mmx quantizer bug for qscale%2==1 (bias was slightly wrong)
Originally committed as revision 895 to svn://svn.ffmpeg.org/ffmpeg/trunk
test simple-idct and ijg int DCT too
do tests with random & sparse matrixes
print systematic error matrixes
Originally committed as revision 892 to svn://svn.ffmpeg.org/ffmpeg/trunk