Respect path boundaries in reference filesystem find

[fhgffy]

ReferenceFileSystem.find() matches raw string prefixes, so find("dataset") includes dataset-other/file, and find("dataset/file") includes dataset/file-extra. A missing path such as data can also return entries from dataset. Both the name list and the detailed result are affected.

This also gives FSMap phantom keys: a mapper rooted at dataset exposes -other/file from its sibling, and its length is inflated. The directory-aware find(..., withdirs=True) path already keeps the correct boundary.

Match either the exact path or a descendant beginning with path + "/". Normalize the input before testing for the root so reference:// continues to list all references. The optimized reference scan remains in place.

Regression tests cover directories, trailing slashes, protocol-qualified paths, exact filenames, missing prefixes, root inputs, both detail modes, and mapper keys/length/values.

Validation on Windows / Python 3.14.3:

• Before the fix: 13 failed / 4 passed in the new regression selection.
• After the fix: all 17 new cases passed.
• Reference, mapping and base specification modules: 359 passed / 121 skipped / 1 xfailed.
• Rebased onto current master 26ebc63: all 17 new cases passed again; neither changed file differs between the original and current bases.
• Ruff 0.14.3 check and format check, and git diff --check, passed.

The complete suite was not run. The related selection includes local FTP fixtures and an expected-failure external FTP test.