This fixes a deadlock bug during MDS shutdown:

- the "signal_handler" thread receives the shutdown signal and invokes
  MDSDaemon::suicide() while holding `mds_lock`

- MDSDaemon::suicide() invokes Beacon::send_and_wait() while still
  holding `mds_lock`

- meanwhile, all "ms_dispatch" threads get stuck waiting for
  `mds_lock`, for example in MDCache::upkeep_main() or
  MDSDaemon::ms_dispatch2()

- Beacon::send_and_wait() waits for a `MSG_MDS_BEACON` packet to be
  dispatched (via `cvar` with a timeout)

At this point, even if a `MSG_MDS_BEACON` packet is received by one of
the worker threads, they will put it in the `DispatchQueue`, but no
dispatcher thread will be able to handle it because they are all
stuck.  The cvar.wait_for() call in Beacon::send_and_wait() will
therefore time out and the `MSG_MDS_BEACON` will never be processed.

The proper solution is to unlock `mds_lock` to avoid the dispatchers
from getting stuck.  And in general, we should be holding a lock
strictly only when it is needed and never do blocking calls while
holding a lock.

Fixes: https://tracker.ceph.com/issues/68760
Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
(cherry picked from commit c0fedb5)