Allow AMI action callback to be reentrant.

Fix AMI module reload deadlock regression from ASTERISK-18479 when it
tried to fix the race between calling an AMI action callback and
unregistering that action.  Refixes ASTERISK-13784 broken by
ASTERISK-17785 change.

Locking the ao2 object guaranteed that there were no active callbacks that
mattered when ast_manager_unregister() was called.  Unfortunately, this
causes the deadlock situation.  The patch stops locking the ao2 object to
allow multiple threads to invoke the callback re-entrantly.  There is no
way to guarantee a module unload will not crash because of an active
callback.  The code attempts to minimize the chance with the registered
flag and the maximum 5 second delay before ast_manager_unregister()
returns.

The trunk version of the patch changes the API to fix the race condition
correctly to prevent the module code from unloading from memory while an
action callback is active.

* Don't hold the lock while calling the AMI action callback.

(closes issue ASTERISK-19487)
Reported by: Philippe Lindheimer

Review: https://reviewboard.asterisk.org/r/1818/
Review: https://reviewboard.asterisk.org/r/1820/
........

Merged revisions 359979 from http://svn.asterisk.org/svn/asterisk/branches/1.8
........

Merged revisions 359980 from http://svn.asterisk.org/svn/asterisk/branches/10


git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@359981 65c4cc65-6c06-0410-ace0-fbb531ad65f3