Kura cache runs on bare-metal regions with gapless rolling deploys

Kura cache regions are now hosted on bare-metal machines instead of cloud VMs. Each bare-metal box has a real NIC and local NVMe, so cache serving is faster and no longer capped by a load balancer. Rolling deploys no longer stall writes: the cache runs a warm standby and the controller fails the cache Service over to a standby that is Ready and already runtime-confirmed serving, so customers only see a brief window of misses on the very newest artifacts, which regenerate on the next fetch.