Stop all application states in dependency order on shutdown

[vsr625]

Previously each mode's stop-fn tore down its own subset of states, and modes were stopped in sorted order. :api-server's stop closed the shared consumer-connection before :worker cancelled its RabbitMQ subscribers and drained in-flight messages. Handlers mid-processing then failed to ack on the closed channel, publishing successfully processed messages to the dead set while the broker requeued them, causing double processing. Shared states (kafka-producers, producer-connection, channel-pool) had the same problem: whichever mode stopped first closed them while other modes' flows were still publishing through them.

Replace the per-mode stop dispatch with a single stop-application-states fn that stops every state in dependency order regardless of modes: stop ingestion (server, streams, batch consumers), then cancel RabbitMQ subscribers and drain in-flight messages, and only then close the connections and producers those handlers publish through. Stopping a never-started state is a no-op in mount, so the sequence is safe for every mode combination.

Each state is stopped with its own mount/stop call because mount orders a multi-state call by namespace load order, not by argument order; separate calls are the only way to guarantee this dependency ordering.

With stop no longer dispatched per mode, drop the now-dead :stop-fn entries from valid-modes-fns. The public stop-* fns are kept as-is for backward compatibility.