use gradients instead of prediction errors as context model
store independant quantization tables for each point
merge contexts with opposit sign

Originally committed as revision 1957 to svn://svn.ffmpeg.org/ffmpeg/trunk