Clean up leftover files before serialization doctests

[lgoettgens]

Resolves #4572.

[aaruni96]

I think this may ultimately not be needed. If you write to a file, and it already exists, you just overwrite it. And if you can't overwrite it, chances are that you can't remove it either (as was the case which motivated this : the file was owned by someone else and the runner process couldn't write to it).

My current concern is what happens when two different tests running on two different runners try to do this to the same file at the same time. Either we make the runners present a different tmp to each run of the test, or, in the test, we use mktemp and write files there (and cleanup after the run).

Thoughts @lgoettgens @fingolfin @benlorenz ?

[lgoettgens]

I think that our CI jobs shouldbe allowed to expect not to share any memory with other concurrently running jobs. In that regard, I think the cleanest solution would be that each runner has its own /tmp directory, as it is the case for the github-hosted runners.

[benlorenz]

I would like to do it with a temporary directory in the doctest, but it should not be user-visible. And so far Documenter doesn't have a cleanup for a doctestsetup (to get back after switching the directory in a setup block).

In this case, I don't think this can trigger a race-condition as long as all runners have the same uid: The files are always identical and are written atomically, i.e. any job can rewrite them and there should always be a file to read from.
(In this case the atomic rewrite failed because of the permissions and then julia used cp && rm as fallback, which also failed)

[fingolfin]

So having JuliaDocs/Documenter.jl#2577 in Documenter would be useful.