fix(monitoring): remove --pid host, consolidate cpu smoke test into c…

[sbates130272]

…ompose

spur-authz blocks --pid=host on all SPUR nodes. The cpu smoke test was using raw docker run with --pid host for node-exporter and nvme-exporter, causing a fatal exit 125. The monitoring compose had the same flag, which has been silently failing in all CI tests where monitoring is best-effort.

• Remove pid: host from node-exporter and nvme-exporter in docker-compose.monitoring.yml; hsa-snoop retains it (kprobe requirement)
• Add exporters-cpu profile covering node/nvme/rdma but not GPU exporters (amdgpu-exporter, hsa-snoop), so the cpu smoke test can use start_monitoring() without attempting to launch GPU-only services
• Switch docker-compose.emulator.yml from bridge+published port to network_mode: host, consistent with the monitoring stack's localhost scraping
• Rewrite spur-monitoring-cpu-smoke.sh SRUN_BODY to use compose + monitoring-lib.sh (start_monitoring/stop_monitoring) instead of reimplementing container orchestration inline with raw docker run
• Add AIC_CPU_SMOKE=1 flag to mon_profile() in monitoring-lib.sh to select the exporters-cpu profile; add aic-vllm-emulator to MON_CONTAINERS

[sbates130272]

/run-ci

[github-actions]

❌ dist-build failed for f7d9244. View logs.

[amd-ivaganev]

IIRC --pid host was required for vllm/lmcache to share GPU memory across containers. Won't that break the purpose of the smoke testing?

[github-actions]

✅ Hardware CI passed for f7d9244 (dist-build + smoke-test + tiny-test).

[sbates130272]

IIRC --pid host was required for vllm/lmcache to share GPU memory across containers. Won't that break the purpose of the smoke testing?

It seems that --pid <container-name> works and is more isolated than using --ipc host.