When a response is very long (e.g. a large playlist > 100k songs),
most of it will end up in the deferred buffers.  Filling the deferred
queue is very expensive currently, because a new buffer is allocated
for every client_write() operation.  This may lead to long delays, and
the client might give up and disconnect meanwhile.  This patch makes
MPD attempt to flush the deferred queue as often as possible, to work
around this problem.  Due to the MPD 0.14 code freeze, we should not
optimize the buffering code now.