Don't lock the VMG_LOADED mutex across global destruction

If this mutex falls back to the OP_REFCNT mutex, and that a coderef is
freed during global destruction, the mutex is taken twice and the process
hangs. As a general rule, it does not seem very safe so we just set
vmg_loaded to zero early and only run the global teardown callback if
it is zero. This ensures that this callback will only be called once.