LiveLesson state manipulation still requires scheduler in some cases

[Mierdin]

antictl and the RPCs it communicates with allow a developer to set up antidote-core in a mode where only the API is running, and the scheduler is effectively disabled. This is immensely useful because it allows people to have a functioning API without having to set up all of the underlying compute bits (i.e. Kubernetes).

In the normal workflow, this works great. Devs can use antictl to create a new livelesson instance. However, deleting that instance doesn't work because the "kill livelesson" RPC doesn't actually delete state - it assumes the schedule is online and will clean up that state when it's done killing the livelesson namespace.

This means that devs need to basically restart antidote-core when they want to recreate state.

Should put more thought into this. a) is this really a problem? and b) is it worth solving, and if so, how? Should we allow the API to delete state without the scheduler confirming it was done properly? Or maybe have this "kill" workflow function with a livelesson ID as a parameter, and that way it doesn't have to look up the livelessson details in internal state, which wouldn't be there at this point. Also should consider if there are any other RPCs that have a similar issue.

Another way to fix this could be to just allow the API to overwrite state - but I can think of some use cases where actually deleting a livelesson is still useful.