This prepares for beforeAll/afterAll hooks to be handled in the same way.
Since we do not know in advance whether a hook will run, we must create
TestResults lazily.
Two bug fixes:
- Do not use the worker that is being shutdown for a new job.
- Report unhandled errors during "expected to fail" tests as
fatal errors.
In some circumstances, dispatcher was waiting for all exisitng jobs
to finish before scheduling a new one. This leads to unneded stalls.
Instead, we can schedule jobs right away, if we have a worker
available.
