This commit ensures that when `TriggerService.start()` is called,
we ensure in the trigger engine implementations that current watches are
removed instead of adding to the existing ones in
`TickerScheduleTriggerEngine.start()`
Two additional minor fixes, where the result remains the same but less code gets executed.
1. If the node is not a data node, we forgot to set the status to
STARTING when watcher is being started. This should not be a big issue,
because a non-data node does not spent a lot of time loading as there
are no watches which need loading.
2. If a new cluster state came in during a reload, we had two checks in
place to abort loading the current one. The first one before we load all
the watches of the local node and the second before watcher is starting
with those new watches. Turned out that the first check was not
returning, which meant we always tried to load all the watches, and then
would fail on the second check. This has been fixed here.
b6f762d131