feat(controller): reconcile dependents when their cluster becomes ready

[bmorton]

Problem

When a TemporalCluster and a dependent object (e.g. TemporalNamespace) are created at the same time, the dependent registers much later than the cluster becomes ready. Two mechanics cause this:

1. No event-driven trigger. The cluster-dependent controllers (TemporalNamespace, TemporalSchedule, TemporalSearchAttribute, TemporalWorkflowRun, TemporalClusterConnection) resolve their target via resolveTarget and, when the cluster isn't Ready, just requeue on a timer. None of them watch the TemporalCluster/TemporalDevServer, so nothing re-triggers a reconcile the moment the cluster flips to Ready.
2. Exponential backoff on transient not-ready. Once the cluster reports Ready, the controllers immediately call the frontend. If it's up but not yet serving RPCs, the call returns a raw error and the controllers return ctrl.Result{}, err — so controller-runtime applies exponential backoff (default max ~16.7 min). The object is "caught in a backoff" and registers long after the cluster is usable.

This also slowed the e2e suites, which sequenced cluster-first-then-dependent to sidestep the delay.

Changes

• Shared helpers (internal/controller/cluster_watch.go): a clusterReadinessChanged predicate (fires on create, generation change, or a Ready-condition transition — ignores routine status writes), a refTargets matcher (same-namespace, empty Kind defaults to TemporalCluster), an isTransientClusterErr classifier (gRPC Unavailable/DeadlineExceeded/Canceled), and clusterUnavailableRequeue = 5s. New reason const ReasonFrontendUnavailable.
• Event-driven readiness: each dependent controller now .Watches TemporalCluster and TemporalDevServer, mapping a changed target to same-namespace CRs referencing it. Dependents reconcile the instant their cluster becomes Ready.
• No backoff on transient not-ready: the four RPC-registering controllers convert transient frontend errors into a short fixed requeue (Ready=False / FrontendUnavailable) instead of erroring. Terminal errors still surface. TemporalClusterConnection already swallowed transient per-peer errors, so it gets the watch only.
• e2e: test/e2e/namespace now applies the cluster and namespace concurrently, proving the fix and trimming wall-clock time.

Testing

• make lint — 0 issues
• make test (unit + envtest) — all pass, no generated drift
• New unit tests for the predicate, matcher, transient classifier, and map function; new envtest spec asserting a transient frontend error yields a short requeue (not a returned error).

No API/CRD schema changes.

Design & plan: docs/superpowers/specs/2026-07-24-cluster-readiness-watch-design.md, docs/superpowers/plans/2026-07-24-cluster-readiness-watch.md.