Client::shutdown() could indefinitely wait when tearing down
MDS sessions if an MDS is unreachable during shutdown, but a
valid session existed, i.e., Client::mount() was successfull.
These failures were initially observed in mgr/volumes tests in
teuthology and lately during rados/mgr selftests where manager
plugins do not respawn as volumes plugin waits for cleaning up
libcephfs handles via Client::shutdown().