Harden concurrency service on restart

[maxfischer2781]

The concurrency service of cobald has been written with just a single run – as a demon – in mind. However, unittests will rerun or recreate the service several times.

Spurious test failures via PyPy were observed on #94. Since PyPy uses a different garbage collection strategy with higher latency, this indicates the concurrency service has implicit dependency on prompt cleanup.

This is not an issue in production, but hints at the code not being robust. The service should be reviewed and possibly rewritten to have proper, explicit lifetime.