Deduplicate coordinator run-loops into a shared actor abstraction

The challenge responder and replica-sync coordinator (and for now the checkpoint coordinator) share the same run-loop scaffolding: a start() that spawns the loop and hands back a command handle, a biased tokio::select! over command/event/interval arms, pause/stop command handling, the events_open disarm for a closed broadcast channel, Lagged → bootstrap-rescan recovery, and the safety-net interval. Only the duty-specific behavior differs.

Raised by @bkontur in https://github.com/paritytech/web3-storage/pull/291#discussion_r3638194942: extract the shared scaffolding into a common trait Coordinator/actor abstraction that owns the tokio/channel/restart handling (including what happens when a coordinator task dies or panics), with per-impl hooks for the relevance filter, event reaction, bootstrap scan, and tick action — without pulling in an actor framework.

Sequencing: this should land after #311 (which removes the checkpoint coordinator, shrinking the surface to two loops) and after #297 (stacked on #291), so the refactor doesn't churn under both.

Relates to #291.

Start by reading #311 and #297, then compare the challenge responder and replica-sync coordinator run loops after those changes land. Identify the shared start, tokio::select!, channel-recovery, pause/stop, restart, and interval behavior, and define hooks for each coordinator's differing actions. Done means both remaining coordinators use one abstraction without an actor framework.