Could we add a shutdown/cleanup mechanism to the testing package?

Background: some tests (notably: integration-style tests) require heavy setup, compiling
things, setting up things on disk, running child processes wired together, etc.

If you just do this once and share it per test, it's unclear when to clean up the /tmp
folder and shut down the processes.

People have resorted to things like TestZZZ for cleanup (which doesn't work well with
-test.run), or z_last_test.go (as seen in net/http), or the hoops jumped through here:

https://github.com/bradfitz/docker/blob/fuse/vfuse/all_test.go

(Look at release and testsToRun and how it's calculated.)

And also hacks in Google's code, even people so far as to fork the testing package.

Or people run a child process that is responsible for cleanup, once it notices its
parent has died. But that's a lot of work and there are portability concerns.

I know a test could still crash mid-way through, so cleanup on shutdown is still best
effort, but when it's a bunch of big processes and big files in /tmp, that's better than
nothing most the time.

With regard to atexit-ish concerns, I'd propose the shutdown mechanism itself be subject
to the same timeout concerns.  If the shutdown doesn't finish in N seconds, the test
exits with failure.