mds: fix mds scatter_writebehind starvation

mds: fix mds scatter_writebehind starvation

scatter_writebehind is called by eval_gather on dirty locks, and
eval_gather is called by wrlock_finish on unstable locks when you
drop the last wrlock...and scatter_writebehind force-takes a wrlock.

This meant that a workload like:
seq 3000|xargs -i mkdir a/b/{} &
mkdir a/c
could cause the mkdir a/c to wait until after the other process
finished because rstats can propagate upwards asynchronously, but
mark the directory dirty synchronously, while the mkdir a/c requires
an actual wrlock in order to modify the rstats.

Signed-off-by: Sage Weil <sage@newdream.net>
Signed-off-by: Greg Farnum <gregory.farnum@dreamhost.com>