feat: continuous re-registration reconciliation (#5)

[as51340]

What

Implements specs/operator-mvp/issues/05-continuous-re-registration.md: the operator's reason to exist over the chart's one-shot setup Job. Registration state a pod loses (rescheduled onto a fresh node, wiped storage) is now restored automatically.

The gap this closes

The registration planner already computed full-diff semantics (from #3). The missing piece was on the controller side: a converged reconcile returned ctrl.Result{} with no requeue. A registration lost on a still-running pod produces no watch event — its StatefulSet is unchanged — so the drift would never be detected. Reconciliation was effectively one-shot.

This adds a resyncInterval (30s) and requeues on it from the converged branch, so a converged cluster is continuously re-observed and drift is re-issued without human action. Leadership stays hands-off: the planner never emits a MAIN promotion when a MAIN already exists.

Changes

• Controller (internal/controller/memgraphcluster_controller.go): periodic resync from the converged branch; docstring updated to describe continuous (not one-shot) reconciliation.

Tests

• Planner (internal/planner/planner_test.go): multiple lost registrations re-issued without a second MAIN (adds the "multiple lost" case; single-lost, converged-no-op, and recovery-with-MAIN-present were already covered).
• Controller envtest (internal/controller/memgraphcluster_controller_test.go): re-register a lost data instance leaving MAIN untouched; re-add a lost coordinator; no-op across repeated resyncs (each reschedules the resync). Updated the bootstrap spec's converged assertion to expect the resync requeue.
• E2E (test/e2e/memgraphcluster_test.go): unregister a data instance on the coordinator leader, confirm it left the view, then assert the operator converges the cluster back to fully registered. Factored the converged-state check into a shared verifyClusterRegistered.

Acceptance criteria

• A data instance whose registration is lost is automatically re-registered on a subsequent reconcile
• A missing coordinator is automatically re-added
• A converged cluster produces zero commands on reconcile (verified no-op)
• No MAIN promotion is ever issued while a MAIN exists, including during recovery
• Planner unit tests cover: single lost registration, multiple lost, converged no-op, recovery with MAIN present
• E2E scenario: wipe one instance's registration state, assert the cluster converges back to fully registered

Verification

make test-unit, controller envtest (go test ./internal/controller/...), and make lint all green locally. E2E runs in CI against the licensed KinD cluster.