Add a way to ensure tests actually test something

[jyn514]

A common problem I run into with tests is that it's hard to verify they actually do anything. Consider for example #102038: this test had no changes, but before that PR did not actually ensure that the issue didn't reoccur.

We could avoid problems like that by adding a new test annotation, // test-fails-with nightly-YYYY-MM-DD. That will make sure the test actually does something and hasn't always passed.

It will be too expensive to run all those versions on each bors run, so we could instead only run it every 6 weeks before a release (perhaps we can fold this into crater how? at least the automation if not the same codebase.)

[jruderman]

Instead of every 6 weeks, how about "when the test changes"?

[jyn514]

@jruderman hmm, yeah this is harder than I thought. "when the test changes" is a sufficient condition for // test-fails-with-nightly. But that's not actually what went wrong in #102038; what went wrong was that it became harder to trigger the underlying bug even though the test was unchanged.

I think to test it properly it would have needed something like 2 revisions, with and without -Z normalize-docs, to make sure it passed without and failed with. But that seems quite hard to generalize. I think test-fails-with-nightly is more likely to be generally applicable, and there "when the test changes" is good enough.