fix(reconciler): exponential crash-loop backoff for whole-pod restarts

[dndungu]

Problem (#54)

Exit-driven restarts had no damping: a pod crashing a few seconds after start was recreated every tick — 128 restarts in ~54 minutes, with each cycle re-running full startup and its external traffic. The existing backoffDelay/retryEligible machinery only gates create failures (podman run errors via StartAttempts); a clean start followed by a crash bypassed it entirely.

Fix

• nextPodBackoff on the Always / OnFailure-retry transitions: delay doubles 10s → 5m cap (Kubernetes CrashLoopBackOff schedule); reconcilePending holds recreation until the deadline.
• Reset after a clean run: a pod that ran ≥10 minutes before exiting starts the schedule over (StartedAt-based).
• The pending event carries the delay (restarting (policy=Always), crash-loop backoff 40s).
• State is reconciler-memory (same trade-off as the One crash-looping container restarts every container in the pod (k8s restarts only the failed one) #46 per-container backoff), swept on terminal state and pod deletion. Deliberately separate from the persisted create-failure backoff — different reset semantics; both documented in the devlog.

Testing

• TestCrashLoop_BackoffDelaysRecreate: fake-clock schedule pinning (no recreate at 5s, recreate at 11s, doubled to 20s, event message carries the delay).
• TestNextPodBackoff_CapsAtMax / _ResetsAfterCleanRun.
• Existing restart tests updated to advance the clock past the backoff instead of asserting immediate recreate.
• Full suite -race green, vet + staticcheck clean.

Live acceptance after deploy: a pod with command: ["sh","-c","exit 1"] under Always must show growing gaps between restarts (10s, 20s, 40s...) instead of one per tick.

Closes #54