Fix glob pattern budgets across exclusions

[JackyHsu-X]

glob.glob('a', exclude='b', limit=1) currently succeeds even though the include and exclusion expand to two patterns. The scanning Glob class starts a new count for each group, so plain patterns and SPLIT expansions can exceed the shared limit. This is separate from the compilation and translation paths fixed in #243.

Keep the expanded-pattern count on the Glob instance across both groups. Only reduce the brace expansion allowance for positive limits, preserving unlimited expansion for zero and negative limits after an earlier pattern.

Add regressions for glob and iglob, string and bytes inputs, include/exclude lists, repeated patterns, BRACE and SPLIT expansion, exact limits, and the implicit NODIR filter. Include a changelog entry.

Validation on Windows / Python 3.12.8:

• New regressions against unchanged production code: 40 failed, 20 passed.
• With the fix: all 60 new cases pass.
• Full suite: 1,342 passed, 167 skipped, 12 subtests passed; 97% combined line/branch coverage. Changed executable lines and decision branches are covered.
• python -m ruff check ., python -m mypy, and git diff --check pass.

The hosted operating-system/Python matrix and documentation build have not been run locally. AI tools assisted with implementation and review; the reproduction and checks were run locally.

[facelessuser]

@gir-bot lgtm