feat(ci-project-cleaner): stale VM cleanup as hourly CronJobs
Stale VM cleanup as CronJobs on runner-managers-gprd-1 (production-engineering#29599): one per ephemeral VM project (56, the r-saas families, gpu, and org), hourly, one project per minute.
Same deletion rules as the ci-project-cleaner pipeline: instances named runner-*, labeled gl_resource_type: ci_ephemeral, older than 3 days. The official gcloud image runs a ConfigMap script, authenticated via Workload Identity as stale-ci-cleaner@gitlab-ci-155816 (config-mgmt!15141). No custom image, no Vault.
The CronJobs start in dry-run mode, logging what they would delete; DRY_RUN flips to false after comparing against the pipeline. The pipeline runs in parallel (cleanup is idempotent) and its schedule is disabled once the CronJobs prove out: its two remaining entries, gitlab-ci-155816 and gitlab-ci-windows, are vestigial (the first has no runner instances, the second's Windows VMs lack the ci_ephemeral label, so the pipeline's identical filter never matched there).
config-mgmt!15141 applies first. Alerting (no successful run for 3h per project) follows in runbooks.