When a parent actor has a fault, child actors should also fault and not keep running

[dulinriley]

As part of our principles of supervision, a child actor should never be able to be orphaned, because then we have no one to deliver any supervision events to.
If the owner of a mesh hits a fault, it won't have a chance to run cleanup and shutdown child actors. This will orphan those actors.
The children should have some way of detecting when their parent is down, and then stopping itself (and all its children recursively).

Something that is key to implementing this right is that it should not rely on the owner or parent to do anything at crash time. These
can be all sorts of crashes, including if the power is yanked on a machine. So we need the children to be able to figure out the parent is
dead on their own.

Some ways to implement this:

• Some periodic message from child actors to their owner to check if they are alive if they get a timely reply back. If the message can't be delivered, it'll go through the undeliverable message handler, which we'll need to make sure also fails the actor
• Piggyback off some other periodic message. For example, an owner should be sending out periodic GetState messages to children to see if they are alive, the children can assume the parent is alive if it gets those messages. Could have a problem if we want to remove that particular periodic message, or if there are other things that can stall sending out messages that aren't actually making the actor dead / unreachable

[dulinriley]

Another important feature of this change: if we implement this, when the root of the tree (aka the client) hits a fault or is Ctrl-C'd we will be able to cleanup everything it spawned.
This is a nice property to have if you leave a cluster up for experiments and have monarch attach to it. Because whenever you exit your client (be it a script, notebook, etc.), it will shut down everything it created. The next time you start your client you can be sure all of the previous actors have been cleaned up.

[johnwhumphreys]

Diff in review still, iterating.