Faster macOS runner provisioning with golden base VMs

macOS runner provisioning now reuses a per-host golden base VM instead of re-downloading the ~40 GB macOS+Xcode image on every job. Each macOS pod is spawned from a local APFS clone of the golden, which removes the network pull and cuts median provision time from around 7 minutes down to seconds. As a side effect, roughly 14 TB/day of redundant image ingest per fleet is eliminated, and steady-state concurrency on the 9-host macOS pool climbs from about 4 running builds to closer to 9, so queues back up less often. Idle hosts also retain their golden for 24 hours, so a burst after a quiet period still clones instantly rather than re-pulling. Pods now also prefer hosts that already hold the relevant golden, keeping cold materializations rare.