Skip in-tree directory links so sdist keeps the real package

[Kayforkind]

Summary

• safe_walk follows directory links and records (st_dev, st_ino) to stop cycles. A junction or symlink whose name sorts before its target (bar → foo) was marking foo's inode as already seen, so the real package was never walked.
• In-tree directory symlinks and Windows junctions are now skipped so the real path is packed. Out-of-tree links are still followed. Cycle detection is unchanged.
• exclude of the alias no longer drops the real package's files.

Fixes #1197.
Fixes #2008.

Test plan

• test_in_tree_dir_link_does_not_hide_real_directory (issue Junctions/symbolic links can cause directories to be missing from source distributions #1197 repro)
• test_excluding_dir_link_keeps_real_directory (Inconsistent/surprising symlink handling during sdist build? #2008 exclude footgun)
• hatch test tests/backend/builders → 469 passed, 10 skipped
• Existing test_infinite_loop_prevention still valid (unix)

AI disclosure

Assisted by Cursor. I directed the change, reviewed inode first-seen-wins vs cycle detection, and verified the tests above. No Co-authored-by LLM trailer.

Made with Cursor

[cjames23]

Do not include issue links or reference numbers in comments. And comments should explain things that are not clearly explainable by reading the code itself.

[cjames23]

Remove issue links from comments

[Kayforkind]

@cjames23 thanks for the master sync. Comment cleanup in �ackend/src/hatchling/builders/utils.py and ests/backend/builders/plugin/test_interface.py is in ef34209 — issue links dropped, comment kept to the inode first-seen-wins trap that isn't obvious from the code. CI is green on the merged HEAD. Ready when you are for another look.