lint_po: ignore unknown-message-flag errors

See merge request !1

This error breaks CI with PO files generated by recent po4a.

I've asked i18nspector upstream to fix this:
https://github.com/jwilk/i18nspector/issues/11

I'll create an issue to revert this once this is fixed in a version
of i18nspector that we can install everywhere relevant.

We don't use them.

This wasn't tidied up in earlier commits to make sure the revert (in the
previous commit) would apply cleanly, but there's of course no need to
keep creating this temporary directory.

[ Due to indentation changes, use -b/-w to double check this commit. ]

This reverts commit c957b5b8.

With the new approach (implemented two commits ago), a cache directory
is created for each i18nspector call. This means that we don't need to
treat the first file specifically (which wasn't working, depending on
which file that was).

As a result, the code can be simplified, getting back to the previous
“treat all files the same way” implementation.

Let's clean things up a little: The previous commit was an incremental
evolution, creating individual cache directories below the previously
global temporary directory. Even if that works fine, that's a little
convoluted, and there's no reason for this indirection: each i18nspector
call can get its own temporary directory anywhere, so simplify how the
temporary directories are created.

Treating the first PO file as special was a nice try but some files
don't trigger the creation of a parser, i.e. don't result in the
creation of a rply cache directory and/or cache file.

That's the case if the first file is wiki/src/getting_started.fr.po for
example, which results in various failure modes, depending on who wins
the race:
 - mkdir on the just-created rply directory;
 - trying to read from the just-created and still-empty cache file,
   which is not valid JSON-wise;
 - trying to read from a being-written cache file, leading to
   inconsistent data, meaning another JSON validity problem.

Therefore, consider the XDG_CACHE_HOME environment variable set at the
top-level to be the top-level directory, below which a specific
subdirectory will be created for each i18nspector call. This should
ensure nobody steps on someone else's toes.

This is just an attempt at a minimal bug fix, follow-up commits will
simplify that code drastically.

When having "lint_po --cached" in one's pre-commit hook, any attempt
to create a commit that does not modify any PO file fails with:

  Traceback (most recent call last):
    File "/home/intrigeri/.home/prv-user/.combined/cyber/amnesia/live/git/submodules/jenkins-tools/slaves/lint_po", line 347, in <module>
      main(logging.getLogger())
    File "/home/intrigeri/.home/prv-user/.combined/cyber/amnesia/live/git/submodules/jenkins-tools/slaves/lint_po", line 327, in main
      chain = [(_check_po_file(next(iterator)),),
  StopIteration

  The po files you're trying to commit contain errors. Please fix them and try again.

This is a regression introduced by c957b5b8,
which b37a38ce fixed already, but only for the
case when --cached is *not* passed.

Note that in --cached mode, we don't display a warning: for details,
see 0da94c64.

Because the plain yaml.load could be easily exploited.
Please read https://msg.pyyaml.org/load for full details.

When there are no files to process on, we should stop the script
directly. So other parts of the script can expect that there are files
to operate on. With the last commit, this is needed, as we execute
next(iterator), that fails if there is no element in the iterator.

As we run check_po for ten files in parallel, we end up in a race
condition about the creation of $XDG_CACHE_HOME/rply. That's why we
check the first file before checking the rest (#17359#note-14).

Without setting XDG_CACHE_HOME to a temporally directory, we end up
using very old ~/.cache/rply.

I understand this error happens because polib has no type hints:
https://mypy.readthedocs.io/en/latest/running_mypy.html#missing-imports

The ":(" smiley unnecessarily conveys bad feelings. Last time I checked,
programs have no feelings :)

Besides, fix typography and make every line of the error message not longer than
80 chars.

Most commits we create don't modify any PO file. It's annoying and confusing to
be told "WARNING: no file to process :(" every time this happens: one may
wonder if they've done something wrong, which is generally not the case.

Thanks in advance.

i18nspector's --language option takes "a 2- or 3-letter ISO 639 language code,
possibly followed by underscore and a 2-letter ISO 3166 territory code"
so we need to strip any "_<SCRIPT>" suffix when passing it the language
we've detected from the PO file name. Otherwise, we get:

   i18nspector: error: invalid language

lang() can return e.g.:

 - fr: language = French, no territory nor script specified
 - fr_CA: language = French, territory = Canada
 - sr_Latn: language = Serbian, script = Latin

Follow-up on 5d4b04c4.

We have this e.g. in greeter.git.

refs: #11081

Counting the number of lines in the output doesn't allow us to
differentiate success (empty output) from failure with one single
error ⇒ let's do it differently.

i18nspector (0.22) unstable; urgency=low

  * Summary of tag changes:
    + Renamed:
      - unusual-plural-forms (from incorrect-plural-forms)

I don't think that patching ikiwiki to avoid this is worth the effort:
we have licensing and copyright info at the repository level.

It comes from the main Tails Git repository at
commit 3d216e76e61514e25e01791ec88e7750c86a90b3.