Ensure tombstones created before kubexit started are read

[RemcodM]

Fixes #8
Replaces #9, this should be a better PR with less unneeded refactoring

This commit changes the Watch function inside the tombstone package to
also emit an fake 'Created' event besides the real fsnotify events. This
initial event is send out immediatly when Watch is called and the watcher
has been setup.

This change allows kubexit to detect tombstones written before kubexit was
started. This prevents possible race conditions as described by #8.

This change on its own introduces a new bug where the tombstone is written
as part of an initial event, but the child process will still start because
child.Start() is being called after the watcher has been setup. To overcome
this issue, the shutdown state of the child is tracked in a new flag, which is
set if ShutdownNow() or ShutdownWithTimeout() is executed.

[RemcodM]

New PR since I decided to make the changes from scratch. I hope you will like this one better 😄

[karlkfi]

If a death dep is already dead, do we still want to wait for the birth deps? That's what your new shutdown flag in child.Start does. It might make more sense to just exit early.

With the new Watch change, any existing tombstones get dealt with synchronously, instead of asynchronously. I think I can make it so ShutdownWithTimeout errors if it's not started, and then fail fast within Watch using an EventHandler that returns an error. If the EventHandler errors asynchronously, the error would just get logged, but if it errors synchronously then Watch can error, which would cause a fatal exit in main.

[tnachen]

Any plan to merge this or @RemcodM's? I'm running on a patched one forever, if this is no longer maintained I guess we just have to fork this.