GeezerVison

`TagObject._set_cache_` split the decoded tag object with `str.splitlines()` and then read the headers by position, taking `lines[3]` as the tagger. `str.splitlines()` also breaks at `\r`, `\v`, `\f`, `\x1c`-`\x1e`, `\x85`, U+2028 and U+2029, but git ends a header only at a line feed. A U+2028 inside the `tag` header of an untrusted tag object therefore shifts every later header by one line in GitPython only. With `tag v1<U+2028>tagger Linus Torvalds <[email protected]> 1700000000 +0000` followed by a real `tagger Mallory <[email protected]> ...` line, `git cat-file tag`, `git show` and `git for-each-ref --format=%(taggername)` report Mallory, while GitPython reported `tag == "v1"` and Linus Torvalds as the tagger. The object passes `git fsck --strict` and a clone with `transfer.fsckObjects=true`, because git only warns `badTagName`.

The fix splits at `"\n"` and drops the trailing empty element, so tags that only use line feeds parse exactly as before. A `\r` or U+2028 inside the message is now kept, as git shows it. The new test in `test/test_refs.py` fails before the change and passes after it. `ruff`, `ruff format --check` and `mypy` are clean. On my machine the full suite has 3 failures (`test_empty_repo`, `test_rebasing`, `test_diff_with_staged_file`), and all 3 also fail without this change.