Fix dedupe scripts leaving duplicate data behind through hardlink siblings - #24

The duplicate map kept only one representative path per inode. If duplicate B had multiple hardlinks (for example, Luminarr and Darkpeers), replacing only Luminarr with a link to A left B's data alive through Darkpeers while the script still reported reclaimed space.

This change retains all known hardlink paths for each physical copy, merges overlapping groups, and groups copies by device and inode. The generated script compares content once per separate copy, creates a replacement hardlink before atomically replacing each included target, and counts each old copy's logical size only when its last hardlink is gone. Failed operations are reported, and partial or completed runs can be safely repeated.

Excluded paths remain untouched, and excluded or unscanned hardlinks prevent a reclamation claim. Filesystem boundaries are respected. The script runs directly on the Unraid host using Bash and standard system tools, with no Python dependency on the host.

Validation:

• 342 backend tests passed, including 18 deduplication cases.
• Tests execute generated scripts on real small files and hardlinks with a PATH containing no Python, covering exclusions, overlapping groups, content differences, link/rename/comparison errors, interruptions, and reruns.
• git diff --check passed.

Limitations:

• Tests were run locally on macOS; the Unraid/GNU tools execution path has not been tested on an Unraid host.
• The Docker build was not run because Docker is unavailable in the local environment.
• SIGKILL may leave .auditorr-dedupe-* temporary directories, although target paths remain present and the script can be rerun.
• Writers must be stopped during deduplication. Reported reclaimed bytes are logical file size, not measured free disk space; open files or snapshots may retain data.