fix: retry hibernate cleanup before deleting a CocoonSet

[CMGS]

Deleting a suspended CocoonSet during a registry probe or DELETE failure previously removed its finalizer anyway. The surviving :hibernate tag could then restore the previous guest state into a new same-named Set, because VM names are stable across Set UIDs. The old Set's cleanup record was already gone, so normal retries and local snapshot GC could not finish the deletion.

Return the hibernate cleanup error and keep the existing finalizer and VM-name record until a later reconcile succeeds. This completes the teardown intent from #5 and #61 using the existing queue and idempotent tag lookup. :latest retention and its best-effort cleanup remain unchanged. Successful reconciliation adds no registry calls, scans, locks, or persisted fields; the production Go diff has zero net lines.

Validation:

• New regression failed on the original code for both probe and DELETE errors, then passed with the fix: cleanup debt survives, and the next reconcile removes the tag and finalizer. A separate case preserves best-effort :latest cleanup.
• GOWORK=off make test: Linux/Darwin vet and the complete native Darwin race/coverage suite passed.
• GOWORK=off make lint: pinned golangci-lint v2.13.2, Linux and Darwin, zero issues.
• GOWORK=off make fmt-check and ASL on Linux/Darwin passed.
• Independent source review and a separate targeted race run found no remaining blocker in this path. No live cluster, registry, or VM test was run.