Fix allcaptures()/allspans() return type in regex stubs

[afonsojanu]

Fixes #566

allcaptures() and allspans() were typed as fixed one-element tuples (tuple[list[AnyStr]] and tuple[list[tuple[int, int]]]), but at runtime they always return one entry per capture group in the pattern, plus the whole match, which is a count that varies from pattern to pattern. A fixed-length-1 tuple is exactly wrong for that shape: a type checker refuses to unpack or index past the first element even though the real return value routinely has more than one.

>>> import regex
>>> regex.match(r"(\w+) (\w+)", "hello world").allcaptures()
(['hello world'], ['hello'], ['world'])

Changed both to the variable-length tuple form, matching what captures() right above them in the same stub already uses for the identical shape.

Added a regression test asserting the corrected static type on both methods, and confirmed it fails with the exact error the issue reports (an assert-type mismatch against the old fixed-length type) when the stub change is reverted.

• python tests/runtests.py --run-stubtest stubs/regex: pre-commit, structure check, Pyright, ty, pyrefly, mypy, stubtest, and both regression tests all pass.

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉

[donbarbos]

Thank you! LGTM, but could you remove the tests?

In typeshed, we only add regression tests for functions and classes which are known to have caused complex problems in the past, or where stubs are difficult to get right. 100% test coverage for typeshed is neither necessary nor desirable, as it would lead to code duplication.

See tests/REGRESSION.md for more information.

And for the record, the issue number refers to the following problem in the library itself: mrabarnett/mrab-regex#566