Postpone the cancellation of authenticator to the very last moment - #1286

Problem: We now cancel all tasks at once (as soon as any task exits, e.g. the exit-flag waiter). However, some tasks require other tasks to properly terminate. For instance, the peering shutdown in the orchestrator requires the authenticator to continue working. With the current approach, the authenticator exits soon, and the peering cannot shut down and de-register itself properly.

Solution: With the new approach, some tasks are made special —the "core" tasks— and we postpone their cancellation to the very last moment, which happens strictly after all the regular ("root") tasks exit, but before the cleanup activities as before.

Backwards compatibility: This way, we keep all the interfaces intact: spawn_tasks and run_tasks() are a part of the public interface (though I regret this move), their change would require semantic v2, which is not planned. Only the internal relations between tasks change, and one task —the authenticator— disappears from the returned list, though remains controlled internally.

Extensibility: When and if needed, some other tasks can be grouped into several layers (exiting priorities) the same way — all hidden from the public interface.

Benefit: a plain and elegant solution to a very specific problem, no over-engineering or over-generalisation (as in the superseded PRs).

Fixes #1033. Supersedes #1110. Supersedes #1282.