Do not share context variables between channels and do not zero them at

decode_frame(). This makes the output much closer to that of the binary decoder.

Originally committed as revision 15244 to svn://svn.ffmpeg.org/ffmpeg/trunk