Move Windows path warning tests (for path containing :) and actually take their results into account

[fingolfin]

This adds back one of the tests removed in #2804 but in a form that's less useless. (See #2804 (comment) for more information about this test.)

However, this will fail CI in the current fork, because the warning (about paths containing :) is actually different on Windows (were our custom code triggers) versus Unix (where a path with a : is fine, but Documenter will complain because the specified files don't actually exist).

So how should we deal with this? Some options:

1. we could just not bother and drop this test (honestly, I think this would be fine, except that it'd drop test coverage a bit)
2. we could extend the check which forbid filenames with a : to trigger independently of the OS
3. we could have two different "expected texts" for this test: one for windows, one for everyone else (similar to how the test/doctests tests allow different versions of a test based on the Julia version)

Thoughts?