Support POSIX character classes in bracket expressions

[dngr2]

The bug

Git's wildmatch supports POSIX character classes inside bracket expressions — [[:digit:]], [[:alpha:]], [[:space:]], [[:upper:]], etc. pathspec does not, for two reasons:

1. The closing-bracket scan stops too early. When scanning [[:digit:]] for the ] that ends the expression, it stops at the ] that closes the inner [:digit:], so the expression is mis-delimited.
2. Python re has no POSIX class syntax. Even delimited, [[:digit:]] is handed to re unchanged, which reads it as a set of the literal characters [ : d i g t (raising a FutureWarning: Possible nested set) — matching nothing sensible.

Result: every POSIX class is broken. Against git check-ignore (the reference), all 12 classes diverge:

>>> import pathspec
>>> s = pathspec.PathSpec.from_lines("gitwildmatch", ["[[:digit:]].txt"])
>>> s.match_file("1.txt")
False          # git ignores it; pathspec does not

A .gitignore rule like [[:upper:]]*.tmp therefore matches files in Git but not in pathspec, so tools built on pathspec (Black, pre-commit, pip, GitPython via GitIgnoreSpec) silently disagree with Git about which files are ignored. This affects both the gitwildmatch and gitignore factories (they share the translation).

The fix

In _translate_segment_glob:

1. Skip [:...:] as a unit when finding the closing bracket, so the inner ] isn't mistaken for the end.
2. Expand each POSIX class to the equivalent ASCII range, matching Git's wildmatch in the C locale ([:digit:] → 0-9, [:alpha:] → A-Za-z, [:punct:] → the punctuation ranges, etc.).

Negated ([![:digit:]], [^[:digit:]]) and mixed ([[:alpha:]0-9], [[:alnum:]_]) forms are handled. Non-class brackets ([a-z], [!a-z], []a]) are unchanged.

Verification

• New test test_15_posix_character_class asserts both the translated regex and match behavior; it fails on the current code and passes with the fix.
• Full suite green: 198 passed, 276 skipped.
• Differentially checked against git check-ignore over all 12 POSIX classes (plus negated/mixed forms) and a 1,500-case fuzz of random bracket/wildcard patterns — 0 POSIX-class divergences remain.

[cpburnz]

Thanks for the pull request. I'll review this along with the similar #128.