LU-3455 mdt: sync when evicting client

LU-3455 mdt: sync when evicting client

If some client is evicted from the MDT, then the underlying OSD
should be synchronized, otherwise the recovery will be failed
for the replay requests were dropped alongside with the client.

in test_20b in replay-single.sh, the occupied disk space will be
released only after DMUs are committed for ZFS, then it should
wait some more time before calling 'df' to get the disk usage info.

This patch is back-ported from the following one:
Lustre-commit: fabd01968d0fac83e4855da5a5dfe8f68c038ceb
Lustre-change: http://review.whamcloud.com/11767

Test-Parameters: alwaysuploadlogs \
envdefinitions=SLOW=yes,ENABLE_QUOTA=yes,ONLY=20b \
mdtfilesystemtype=zfs mdsfilesystemtype=zfs ostfilesystemtype=zfs \
clientcount=4 osscount=2 mdscount=2 \
mdssizegb=10 ostsizegb=10 austeroptions=-R failover=true iscsi=1 \
testlist=replay-single,replay-single,replay-single,replay-single,replay-single

Signed-off-by: Hongchao Zhang <hongchao.zhang@intel.com>
Change-Id: Ia0416e57d009f2fa00c880c6ea41f9d0bd60fef0
Reviewed-on: http://review.whamcloud.com/13231
Tested-by: Jenkins
Reviewed-by: Nathaniel Clark <nathaniel.l.clark@intel.com>
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Bob Glossman <bob.glossman@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>