fix(cp): sweeper task has no panic supervision or liveness signal

[chaodu-agent]

Description

crates/openab-cp/src/main.rs spawns run_sweeper fire-and-forget; the JoinHandle is only used for the final abort(). If the sweeper panics, the CP keeps accepting connections and delegations while lease expiry and deadline sweeping silently stop — leases never expire, deadlines never fire, and no health/readiness endpoint exposes the failure. (#1469 round-4 F11.)

Same failure class as openab's own session-pool reaper bug fixed in #1457: a background loop whose silent death converts a maintenance function into unbounded state growth.

Steps to Reproduce

1. Run openab-cp and force a panic inside run_sweeper (e.g. via a test-only injection).
2. Observe the process stays up, /health still returns ok, registrations/delegations continue.
3. Leases of disconnected agents never expire; deadline-overdue delegations never resolve.

Expected Behavior

Sweeper death is either fatal (process exits, orchestrator restarts it) or recovered (task restarted with backoff), and /health reflects sweeper liveness either way.