The test can randomly fail for two reasons:
1) "cancel" may be called on a job that has been scheduled but was not
running yet. Then the result status is null instead of CANCEL.
2) "interrupt" may be called before "join" is called, then the interrupt
is not processed
The changes ensure that "interrupt" is called after "join" blocked and
that the job is terminated after the interrupt occurred, no matter in
which state it was before.