Fix ensure_empty_dir leaving directory symlinks behind

[Str0k]

ensure_empty_dir in pyproject_api/_util.py used rmtree(sub_path, ignore_errors=True) for every entry that is_dir(). When an entry is a symlink pointing to a directory, rmtree raises OSError('Cannot call rmtree on a symbolic link'), which ignore_errors=True swallows, so the symlink is silently left in place and the directory is not emptied. The callers, e.g. _check_metadata_dir in _frontend.py:348, rely on it to 'start with fresh' before prepare_metadata_for_build_wheel/editable, so a stale symlink in the metadata directory survives and the subsequent mkdir(exist_ok=True) does not restore a truly fresh state. After the fix the rmtree call is guarded with 'and not sub_path.is_symlink()', so symlinks to directories are removed via unlink like any other non-directory entry, without following or traversing the link target.

shutil.rmtree documents that calling it on a symlink raises OSError; combined with ignore_errors=True this turned a hard error into a silent no-op, violating the function's documented contract of leaving the path empty. Trigger: run prepare_metadata_for_build_editable twice with the same metadata_directory that contains a symlink to a directory; on the second run the directory is not emptied. The fix is one guard clause in the existing branch, keeping rmtree for real directories and unlink for everything else, including symlinks; the regression test asserts both that the link is removed and that its target's contents are untouched.

Validation:

• Regression on unchanged base: 1 assertion failure(s); with patch: 1 tests, exit 0.
• Full suite: base 61 tests (exit 1); patch 61 tests (exit 0).

AI assistance: implementation and independent review used Hermes with GLM 5.3. Test evidence was reproduced in clean checkouts. This does not represent a human review.
An additional model reviewed the supplied patch and evidence without executing tools.