Fix symlink prefixes for multiple globstar captures

[shkyyy18]

Reproduction

With GLOBSTAR | REALPATH, the pattern **/literal/**/file.txt can incorrectly match outer/literal/link/leaf/final/file.txt when link is a symlink and FOLLOW is disabled. The second globstar checks outer/link instead of outer/literal/link. Conversely, a symlink at unrelated outer/link incorrectly rejects the real path.

_fs_match initializes base only for the first nonempty capture, losing literal components between later captures.

Fix

Initialize the filesystem prefix from each capture's actual start offset. No changes to parsing, symlink caching, or FOLLOW behavior.

Regression tests and validation

• Eight synthetic filesystem cases (two/three globstars, text/bytes paths, actual-path vs unrelated-path symlink responses) all fail before the source fix and pass afterward. Only islink responses are mocked; paths/files and matching are real. Each also checks the FOLLOW control.
• Added an actual symlink regression to the existing symlink test class. Windows here cannot create symlinks (WinError 1314), so this and existing platform-dependent cases are skipped locally, not claimed as executed.
• Python 3.12.10 / Windows: baseline 1243 passed, 166 skipped; after 1251 passed, 167 skipped, 98% coverage.
• python -m mypy: success, 10 source files.
• python -m ruff check .: passed.
• Full tox/Python/platform matrix and docs build not run locally. Remote CI subject to upstream approval/results.

Read contributing guidance; checked current main and existing issues/PRs. AI-assisted investigation, implementation and synthetic tests; listed validation was executed locally. No personal data or unrelated refactoring.

[facelessuser]

@shkyyy18 This is a good catch. Please see comments and address them.

[shkyyy18]

Addressed both review comments in fb97480: moved the eight regression cases into tests/test_globmatch.py and backticked the globstar term in the real-symlink test docstring. The source fix is unchanged. Local Python 3.12 verification: 1251 passed, 167 skipped, 98% coverage; Ruff and mypy passed. The local test run uses a workspace-local temporary directory because the default Windows temp directory was inaccessible. I could not run pyspelling locally because hunspell/aspell is unavailable; the updated docstring uses the inline-code exclusion already configured in .pyspelling.yml. The new remote CI run is not yet reported as passed.

[facelessuser]

@gir-bot lgtm