Problem
=======
When 'UNINSTALL PLUGIN' and 'START GROUP_REPLICATION' command runs
in parallel, there is a crash caused by trying to destroy a locked
mutex.

Analysis
========
The problem can be caused by the following scenario:
Step1: The 'UNINSTALL PLUGIN' thread holds the plugin_running_mutex
       to stop GROUP_REPLICATION, then switch the execution to
       'START GROUP_REPLICATION' thread.
Step2: The 'START GROUP_REPLICATION' thread waits before acquiring
       the plugin_running_mutex, then switch the execution to
       'UNINSTALL PLUGIN' thread.
Step3: The 'UNINSTALL PLUGIN' thread releases the plugin_running_mutex
       after stopping GROUP_REPLICATION, then switch the execution to
       'START GROUP_REPLICATION' thread.
Step4: The 'START GROUP_REPLICATION' thread holds the
       plugin_running_mutex to start GROUP_REPLICATION, then switch
       the execution to 'UNINSTALL PLUGIN' thread.
Step5. The 'UNINSTALL PLUGIN' thread destroys the
       plugin_running_mutex locked by the
       'START GROUP_REPLICATION' thread.
The 'UNINSTALL PLUGIN' thread calls group_replication_cleanup() to
stop GROUP_REPLICATION and set group_replication_handler to NULL. The
'START GROUP_REPLICATION' thread calls group_replication_start() to
start GROUP_REPLICATION if group_replication_handler is not NULL. To
avoid above problem, the executions of group_replication_cleanup()
and group_replication_start() must be mutual exclusive.

Fix
===
To fix the problem, introduce a new global mutual lock to
make the executions of group replication plugin handler
function accessors be mutual exclusive.