gh-156002: Keep reading through third-party zipfile decompressors

[rasmusfaber]

Follow-up to GH-156003, as discussed in #156002 (comment).

The change to bound decompression made ZipExtFile._read1() call decompress(data, max_length) on non-deflate decompressors and check needs_input. Many third-party decompressors (zipfile-zstd, zipfile-deflate64, and others) that are installed by replacing _get_decompressor() (zipfile-zstd, zipfile-deflate64, ...) do not support those, and end up failing with AttributeError: ... has no attribute '_needs_input'.

This PR:

• gives _decompressor_needs_input() a default argument for decompressors that report neither needs_input nor _needs_input (no three-state return);
• decides once per decompressor instance, when ZipExtFile creates it (construction and seek()), whether it takes the bounded path: only decompressors that report needs_input do, which is all of the stdlib ones. Anything else is read unbounded, exactly as before gh-156002: Bound zipfile decompression for bzip2/LZMA/Zstandard #156003;
• adds a test with a fake third-party compression method whose decompressor has a one-argument decompress() and no needs_input; it fails with the AttributeError above without the fix.

Verified with the real packages by applying the same three hunks to the 3.13 backport (#156738): a Deflate64 archive written by 7-Zip reads through zipfile-deflate64, and a Zstandard member reads through zipfile-zstd, on Python 3.13.15 where both currently fail. test_zipfile, test_zstd, test_zipimport, test_zipapp and test_shutil pass on a debug build of main with _zstd enabled; the *BoundedDecompressTests still pass for every stdlib method.

Should ride along with the open backports #156737–#156741.

Claude Fable 5.1 and GPT-6 Astra both reviewed and gave input, as well as drafted PR description and commit messages.

• Issue: Memory exhaustion via crafted zip file #156002

[python-cla-bot]

All commit authors signed the Contributor License Agreement.

[encukou]

Thanks! I'll push to this PR directly, hope that's fine.

• The comments should be clear about monkey-patching not being supported.
• Adding _needs_input (and not needs_input) to LZMADecompressor was not 100% necessary (see the review; this PR pushes toward the public name to simplify implementation.
• I don't like using the existence of needs_input as a proxy for decompress taking two arguments; I prefer catching the TypeError.
• The fallback to one-argument decompress() is a maintenance burden; let's schedule removal. (The DeprecationWarning should only be for 3.16, not in the backports.)
• In the tests, check that the data is actually "compressed" (via swapcase()).

Could you check if this still works for you?

[rasmusfaber]

Thanks! I'll push to this PR directly, hope that's fine.

Of course. I fixed a small typo that caused CI to fail, and a few other tiny nits.

Could you check if this still works for you?

Yes. This still works fine. I will see if I can get fixes merged in the third-party decompressors so they will continue to work when the fallback is removed (and be safe against CVE-2026-15310) even without a proper public API.