This is more efficient as in practice the check is redundant most of the
time. Callers which do not know if enough data is available have to check
it with av_fifo_size(). Doing the check in *read() means the caller has
no choice to skip the check when its known to be redundant.
Also the return value was never documented in a public header so
changing it should not break the API. Besides this fixes the case where
read() failed on a 100% full fifo.

Originally committed as revision 13404 to svn://svn.ffmpeg.org/ffmpeg/trunk