This turns the output in reproducible-builds/diffoscope#245 from 32MB ->
26MB uncompressed (although almost neglible gain when gzipped, of course).

Gbp-Dch: ignore

Actually don't emit "Unable to stat file" debug messages at all; we have entirely-artificial directory "entries" such as ELF sections which, of course, will never exist as files.

Don't emit a "Returning a FooContainer" - we already emit "Instantiating a FooContainer" log message.

Gbp-Dch: ignore

Don't call difflib.Differ.compare with very large inputs; it is at least O(n^2) and makes diffoscope (appear to) hang. (Closes: reproducible-builds/diffoscope#240)

Gbp-Dch: ignore

Gbp-Dch: ignore

Strip newlines when determining Black version to avoid "requires black >= 20.8b1 (18.9b0\n detected)" in test output.

Fix tempfile/weakref handling in Python 3.7 (ie. Debian buster). (Closes: reproducible-builds/diffoscope#239)

Gbp-Dch: ignore

Mark test_apk.py::test_android_manifest as being allowed to fail for now. (Re: reproducible-builds/diffoscope#237)

Don't rely on magic.Magic(...) to have an identical API between file's magic.py and PyPI's python-magic library. (Closes: reproducible-builds/diffoscope#238)

May break with device-tree-compiler in buster.

This reverts commit cbb229f3.

This can happen during (for example) tests that call `main(...)`.

Ensure we cleanup our temporary directory by avoiding confusion between the TemporaryDirectory instance and underlying directory. (Closes: #981123)

We were calling shutil.rmtree on the TemporaryDirectory instance, rather
than its `.name` attribute. Also add a comment about why we can't store
the directory name and we need to keep the instance around.

Gbp-Dch: ignore