Kura cache moves to bare-metal with gapless rolling deploys

Kura cache regions now run on bare-metal hosts with local NVMe, replacing cloud VMs and load-balancer-fronted gateways. Cross-region replication now goes pod-to-pod over the host NIC with no load balancer in the path. Each region runs a warm standby that the controller promotes during deploys, so rolling upgrades no longer drop cache traffic: the cutover is gapless, and the only cost is a brief window of misses on the very newest artifacts, which regenerate. Region IDs, public hostnames, and the CLI surface are unchanged.