Propagates an instance's bans across a cluster of nodes.

By default all of an instance's state is per node, which works well behind a load balancer: rates and limits are approximate per node anyway. Bans are different: a client banned on one node should be banned everywhere. With

config :my_app, Limen, cluster: [enabled: true]

the instance runs this process in a dedicated :pg scope (Erlang's distributed process groups), named after the instance unless :scope says otherwise, so instances of the same name on different nodes share their bans and other instances never see them. Bans and unbans made on a node (by policies or with Limen.ban/4) are queued in an ETS outbox on the request path, and this process broadcasts the queue to the other nodes every :interval milliseconds. Received bans keep their absolute expiry, so node clocks should be roughly in sync (NTP). Counters stay local.

Nothing on the request path talks to this process; it only reads and writes ETS.