This reduces the amount of memcpy() by using pointers to overlap samples for
successive blocks rather than copying.

Originally committed as revision 25986 to svn://svn.ffmpeg.org/ffmpeg/trunk