Perform the DC prediction reversal immediately after decoding all of
the DC coefficients. This has a greater probability of leveraging the
coefficients while they are still cached.
When testing with the Big Buck Bunny 1080p video, I consistently saw
improvements of 500k-600k dezicycles per run (through
reverse_dc_prediction()) thanks to this move.
Originally committed as revision 19966 to svn://svn.ffmpeg.org/ffmpeg/trunk