Divide first audio buffer chunk into atomary bufffers.

This slightly simplifies decoder and removes potential
audio buffer overrun.

Originally committed as revision 7121 to svn://svn.ffmpeg.org/ffmpeg/trunk