LU-12080 lnet: clean mt_eqh properly

LU-12080 lnet: clean mt_eqh properly

There is a scenario where you have a peer on your recovery queue
that's down. So you keep pinging it, but every ping times out
after 10 seconds. In the middle of these 10 seconds you perform a
shutdown. First you try to do the rsp_tracker_clean. It goes through
and calls MDUnlink on the MD related to that ping. But because the
message has a ref count on the MD, it doesn't go away. The MD gets
zombied. And just waits for lnet_md_unlink to be called in
lnet_finalize(). Then you hit clean_peer_ni_recovery. We see the peer
on the queue, we try to call Unlink on it, but when we lookup the
MD using lnet_handle2md() we can't find it. Afterwards we try to clean
up the EQ and it asserts. Even if we remove the assert we end up with
a resource leak since the EQ is not actually freed since we won't call
LNetEQFree() again.

The solution is to pull the EQ create in the LNetNIInit() and deletion
happens in lnet_unprepare. By this point all the remaining messages
would've been finalized and all references on the EQ are gone,
allowing us to clean it up properly

Lustre-change: https://review.whamcloud.com/34477
Lustre-commit: 1065c8888e96fef9e98676bd3a71b46f7910b085

Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I7fd6018ee2e57f82c649fc3658352e89a4309986
Reviewed-by: Olaf Weber <olaf.weber@hpe.com>
Reviewed-by: Chris Horn <hornc@cray.com>
Signed-off-by: Minh Diep <mdiep@whamcloud.com>
Reviewed-on: https://review.whamcloud.com/36029
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>