windows: always work around EINVAL in case of broken pipe for stdout / stderr In 29a905fe23ae, I missed the fact that the `winstdout` class works around two unrelated bugs (size limit when writing to consoles and EINVAL in case of broken pipe) and that the latter bug happens even when no console is involved. When writing a test for this, I realized that the same problem applies to stderr, so I applied the workaround for EINVAL to both stdout and stderr. The size limit is worked around in the same case as before (consoles on Windows on Python 2). For that, I changed the `winstdout` class.

tests: make pipes / PTYs non-inheritable in test-stdio.py A following patch requires that to test closing the receiving end of the pipe / PTYs. Even for existing tests, it might be safer to make the lifetime of the pipes / PTYs as short as possible.

util: enhance `nullcontextmanager` to be able to return __enter__ result This makes its functionality and signature equivalent to Python 3.7’s contextlib.nullcontext().

windows: augment docstring of `winstdout` class

discovery: change users of `outgoing.missingheads` to `outgoing.ancestorsof` The attribute `missingheads` was recently renamed to `ancestorsof`, as it, despite the old name, doesn’t contain the missing heads but the changesets that were requested (including ancestors) for the outgoing operation. Changing all the users enables to print a warning if the old name is used. There is a good chance that some of the users are buggy because of the old name. Changing them to use the new name makes it more obvious that they are buggy. All users need to be reviewed for bugs. When sending patches for fixing them, the change will be more obvious without having to explain again and again the discrepancy of the old attribute name and what it actually contained.

discovery: fix docstring of `outgoing` class Also, introduce a more correct name `ancestorsof` for what was named `missingheads` before. For now, we just forward `ancestorsof` to `missingheads` until all users are changed. There were some mistakes in the old docstring / name: * `missingheads` (new name: `ancestorsof`) contains the revs whose ancestors are included in the outgoing operation. It may contain non-head revs and revs which are already on the remote, so the name "missingheads" is wrong in two ways. * `missing` contains only ancestors of `missingheads`, so not *all nodes* present in local but not in remote. * `common` might not contain all common revs, e.g. not some that are not an ancestor of `missingheads`. It seems like the misleading name have fostered an actual bug (issue6372), where `outgoing.missingheads` was used assuming that it contains the heads of the missing changesets.

discovery: weaken claim about returned common heads if ancestorsof are given As the test case shows, the claim is not true in general.