only avctx->{width,height} and don't touch coded_{width,height} when parsing
them. This fixes the case when coded and display dimensions differ by more
than one macroblock.

Originally committed as revision 18665 to svn://svn.ffmpeg.org/ffmpeg/trunk