Add atomic write_env_usage test. Fix write_env_usage

[IanButterworth]

#2661 intended to fix #2633 but introduced different failure modes and was evidently missing tests.

Update: This PR

• Adds a test for atomic write_env_usage across processes
• Fixes write_env_usage
• Uncovered a similar bug updating the registry from different processes on windows Registry downloads not atomic across processes: IOError: unlink("registries\\General.tar.gz"): resource busy or locked #2746

[IanButterworth]

Might not be ready for review yet. Just trying to get every platform failing first

[DilumAluthge]

Ooof. Sorry, I jumped the gun!

[IanButterworth]

I think this is in good shape now. For some reason some ubuntu instances survive 1000 iterations, but the rest fail after 1 or 2

[IanButterworth]

Ok, the test seems stable across all platforms and is now marked as @test_broken.

Should I merge this as-is, or keep this test on the shelf if someone comes up with a fix?

cc. @KristofferC @StefanKarpinski

Edit: Maybe it should stay on the shelf. Those windows hangs aren't good

[StefanKarpinski]

I'm generally in favor of merging broken tests rather than leaving PRs around and then opening an issue to track fixing the broken tests. Of course, unless the tests are reliably broken and don't hang, that's a bad idea, but if the broken tests aren't flaky, then I'd merge them.

[IanButterworth]

Ok. To be clear, this PR now adds the test and actually fixes write_env_usage, but uncovered a similar registry bug #2746

So that would be the test failure that would be merged with this.

[fredrikekre]

Can we vendor https://github.com/vtjnash/Pidfile.jl instead?

[IanButterworth]

Seems like a good idea to me and it's not much code. @vtjnash have you ever considered absorbing Pidfile into Base?

[vtjnash]

Sure, that seems great

[IanButterworth]

Superseded by #2793