Runner-job Metrics tab now shows in-VM machine stats

The runner-job Metrics tab now populates with real CPU, memory, network and disk samples. Metrics sampling moved from the runners-controller into the runner VM itself, because the controller’s kubelet/proxy source was unavailable on the macOS Tart fleet (which doesn’t expose the cAdvisor Summary API) and on Linux kata nodes. macOS Tart runners sample whole-VM stats for the job’s full lifetime; Linux adds a dedicated metrics sidecar next to dind. Samples stop cleanly at job end and fail-open on any error so they never block the build.