[Amy (amyxu-08)]

Part of the optimizations outlined in Investigating Substrate Image Pull Speed

Unpacking a large image layer on a cache miss goes through redundant path lookups through *os.Root and uses per-file buffer allocations. For each regular file, unpackLayer walked the path six times (MkdirAll, Lstat in replaceExisting, OpenFile, Lchown, Chmod, and Chtimes) and allocated a fresh 32 KiB buffer inside io.Copy when (*os.File).ReadFrom fell back to genericReadFrom.

• Apply owner and mode on the open *os.File before closing it via tarutil.Fchown and setOpenFileMode.
• Track directories created during the unpack in existingDirs so root.MkdirAll runs once per parent directory, and clear existingDirs in replaceExisting.
• Create regular files with O_CREATE|O_WRONLY|O_EXCL and call replaceExisting only on os.ErrExist.
• Wrap outFile in writerOnly and pass a single 32 KiB copyBuf to io.CopyBuffer.

AI use: AI-assisted, I reviewed and tested it (go test -race -count=3 -shuffle=on ./internal/imagecache andhack/run-e2e.sh against dev cluster).

• Tests pass
• Docs updated (none applicable)