[3.2] Fix race condition on trace_api_plugin shutdown

[heifner]

Failure https://github.com/eosnetworkfoundation/mandel/runs/7143297502?check_suite_focus=true pointed to an issue in shutdown of the trace_api_plugin.

slice_directory::stop_maintenance_thread() sets atomic bool _maintenance_shutdown = true and calls notify_one() on the _maintenance_condition condition variable. This is a race condition because it doesn't acquire the condition variable mutex setting up the possibility that _maintenance_shutdown can be set to true and notify_one called after the while check in slice_directory::start_maintenance_thread but before the wait() causing it to wait forever. This then blocks the slice_directory::stop_maintenance_thread() call of join() on the main thread, blocking the shutdown of all other plugins.

Changed the logic to correctly use a mutex and condition variable around _best_known_lib & _maintenance_shutdown. Also made these two variables non-atomic since they are now only accessed via the mutex. Also release the mutex before running the run_maintenance_tasks() which appeared to be the original intention with the use of the local variables.