Non-clustered NETCONF topology leaks schema registrations

This issue is correctly flagged by SpotBugs and has escaped attention for quite some time. When we are creating the context for a device, we add yanglib if it is configured for that particular device.

If there are multiple devices pointing to the same url, we end up registering for each of them, and if they point to the same schema directory we end up with quadratic number of registrations.

At the very least we should be tearing down the registrations when the device goes away.