Fix delete_original crash for regular checkpoint files

[Nikhi00718]

Summary

delete_original=True currently crashes while deleting a regular checkpoint file. split_and_save_layers() passes a Path, but remove_real_and_linked_file() compares os.path.realpath()'s string result with that Path, assumes the file is linked, and removes the same file twice.

This patch checks os.path.islink() directly, so regular files are removed once and symlink targets are removed only for actual symlinks.

Fixes #297.

Verification

• Added regression coverage for absolute Path, relative Path, absolute string, and symlink inputs.
• 47 passed, 1 skipped in the focused splitter/LoRA/test suite; the skipped test requires Windows symlink permission.
• python -m compileall -q air_llm eval examples rlhf scripts training anima_100k
• Verified a tiny end-to-end split_and_save_layers(..., delete_original=True) run.

Please review when you have a chance, @lyogavin.

[Hotragn]

Verified this one works — I ran it against the real call shape (split_and_save_layers passes a Path) and a relative-string call, and both delete exactly once with no crash. os.path.islink() is the right question to ask; the old realpath(to_delete) != to_delete comparison was always true for a Path because realpath() returns a str.

Just so it's visible to whoever triages: there are four open PRs on this same helper.

PR	opened	approach	result on my test
#303	2026-07-19	os.path.islink()	PASS
#337	2026-08-07	islink + os.fspath + swallows FileNotFoundError	PASS
#362 (this)	2026-09-09	os.path.islink()	PASS
#333	2026-08-05	only initialises targetpath = None	still FileNotFoundError

Your diff is logically identical to #303, which landed first and also ships a test. Nothing wrong with your version — just flagging the overlap so the maintainer doesn't review the same change three times, and so you don't spend more effort on something already covered. #337 is the only one that differs meaningfully: it normalises with os.fspath and tolerates an already-missing file, which matters if the HF cache blob is shared between snapshots.

(I couldn't exercise the symlink path — no symlink permission on this box without elevation — so that part is unverified by me.)