Reduce decoder latency.

Reduce decoder latency.

If large/many FEC slices are used, the decoder gets large amounts of
data in one chunk. Currently it decodes as much as it can which may
take several 100 milliseconds on slow machines -- enough to cause
buffer underruns for the alsa writer.

This patch teaches the decoders to convert only a small amount of data
in one go in order to give the other buffer tree nodes a chance to run.