Nullify pointer to ClientListener in CustomClientInfo in destructor

[sloretz]

This mitigates (does not fix) a race condition when one thread calls rmw_wait() and another calls rmw_destroy_client(). When rmw_wait() wakes up it calls CustomClientInfo->listener_, but both the instance of CustomClientInfo and the object pointed to by listener_ have been destroyed. This nulls the field listener_ and has rmw_wait() check this field before using it to reduce the likelyhood of a crash.

I observed this race condition while debugging a CI test failure in ros2/rclcpp#488. The symptom seen is a SIGABRT on OSX when trying to lock ClientListener.internalMutex_.

I'm not sure what a real fix would look like. rmw_wait() needs a way to tell the client it is holding onto has been destroyed. The other entities appear to be using the same pattern, so I suspect they suffer from similar race conditions too. (ticket to follow)

[dirk-thomas]

I would suggest to wait with merging this until #203 has been merged or it has been decided to delay it for now.

[mikaelarguedas]

My understanding is that #203 will not be integrated before feature freeze so we can move these race condition mitigation forward for the release.

@sloretz is this PR ready for review ?

[dirk-thomas]

If #203 doesn't necessarily change the default behavior and we don't have concerns about regressions it is imo still on the list to possibly be merged.

[sloretz]

Closing this since there have been enough changes to cause conflicts and it only mitigated an issue.