Summary from @kylecarbs:

1. When an agent disconnects, we can't close all Coder server -> workspace connections because it could be intermittent. e.g. if you have multiple Coder replicas, one shuts off due to an upgrade but that was where the agent was connected, we don't want to lose all other client connections.
2. The agent should have an "instance ID" associated with itself to identify when a new instance popped up. When it does, we should publish to all replicas to close connections that are for the old identifier.