Faster macOS runner provisioning by cloning from a local golden base

macOS runner provisioning now clones each VM from a per-host, per-image golden base (a local APFS clonefile, no network) instead of re-downloading the ~40 GB macOS+Xcode image on every single-shot recycle. Median pod provision delay drops from around 7 minutes to seconds, roughly 14 TB/day of redundant image pulls are eliminated, and steady-state concurrent macOS builds on the 9-host fleet increases from around 4 to around 9. Unreferenced goldens are kept locally on each host for 24h so the next job burst clones instantly, and Pods are steered via soft node affinity toward hosts that already hold the matching golden.