More reliable test infrastructure

There are too many @RandomlyFails tests. SemaphoreStep should be used more consistently in place of WatchYourStep and waiting for the execution to suspend. (Already prototyped in WorkflowTest.env.)

Also any build logs should be streamed immediately to stderr, rather than forcing the test to include the current log in every assertion message observed to fail. (In combination with SemaphoreStep, this should also more reliably flush recent output: WorkflowRun flushes logs when new steps are run.)