…Set scale-to-zero

The OT-CONTAINER-KIT redis-operator continuously reconciles its managed
StatefulSets back to the desired replica count. When Sablier scales a
redis-operator-owned StatefulSet to zero the operator immediately restores
the replica count, making scale-to-zero ineffective.

Before scaling a StatefulSet to zero, check whether it is controlled by a
Redis CR (via ownerReferences). If so, set the redis.opstreelabs.in/skip-reconcile
annotation on the owning Redis CR to pause the operator's reconciliation loop.
After scaling back up, remove the annotation to restore normal operation.

The annotation patch is best-effort: if it fails (e.g. the CRD is absent or
the dynamic client is unconfigured) a warning is logged and the scale proceeds
unchanged, preserving existing behaviour for non-redis-operator StatefulSets.

fix(kubernetes): address code review findings on redis-operator skip-reconcile

Correctness:
- Annotation no longer leaks when scale fails: a cleanup defer is registered
  immediately after setting skip-reconcile=true, and clears it if p.scale()
  returns an error.
- Annotation is only set on the scale-to-zero path, not on scale-mode stops
  (sc.Idle.Replicas >= 1), so the operator is not paused while the StatefulSet
  remains non-zero.
- Deferred annotation removal in InstanceStart now uses context.WithoutCancel
  so a cancelled request context cannot leave skip-reconcile set after a
  successful scale-up.
- Both InstanceStop and InstanceStart log a warning when the StatefulSet fetch
  fails rather than silently skipping annotation management.

Quality:
- redisOperatorOwner matches on API group only (not the full version string),
  so the fix continues to work if the operator promotes from v1beta2 to v1.
  The now-redundant redisOperatorAPIVersion constant is removed.
- The redundant StatefulSet GET on the stop path is eliminated by moving the
  annotation block to after getWorkloadLabels (which already fetches the
  StatefulSet), placing it immediately before p.scale(ctx, parsed, 0).
- New unit tests in statefulset_redis_operator_test.go cover redisOperatorOwner
  (including forward-compat with v1), apiVersionGroup, setRedisOperatorSkipReconcile
  set/clear, no-op for plain StatefulSets, stop annotation set/cleanup-on-failure,
  and start annotation removal after successful scale.

Installs a minimal Redis CRD into the shared k3s cluster (following the
same pattern as the CNPG integration test) and creates a Redis CR plus a
companion StatefulSet whose ownerReference points to that CR, simulating
what the redis-operator would produce.

The three sub-tests verify Sablier's behavior against a real API server:
- stop: InstanceStop sets skip-reconcile on the Redis CR and scales the
  StatefulSet to 0
- inspect: InstanceInspect correctly reports the stopped state
- start: InstanceStart scales back to 1 and clears the annotation

No operator binary is run — the tests validate Sablier's own API
interactions with real Kubernetes objects, not operator behavior.