This PR restructures the concurrency runners to be more robust. Major changes include:

• Reduce public features of private types.
  • Remove paranoid thread concurrency protections.
• Switch runner core loops to asyncio from threading
  • This should be more efficient due to less thread switching and waits.
• Move queueing-before-start mechanism to meta runner.
  • This simplifies the actual runners.
  • Runners now use an async queue internally which avoids busy looping.
• Wrap main thread asyncio loop instead of creating a new loop
  • I have no idea how that even worked before w.r.t Run asyncio loop in main thread #8.

Unittests are down from 15-20s to 2-3s now (which doesn't mean that cobald itself is that much faster.)

Note to reviewers:

The public interface of register_payload/run_payload has stayed intact but the parts used by cobald itself has changed considerably. Since the architecture changed (from threading to asyncio) it is probably not helpful to look just at the diff directly.

As a recap of the current architecture:

• The user facing entry point is ServiceRunner, which has not changed other than some documentation.
• The next step is MetaRunner, which now runs an asyncio loop instead of relying on ThreadRunner for concurrency.
  • MetaRunner starts and monitors several BaseRunner instances concurrently.
• The BaseRunner is the template for actual implementations.