fix: restart containerized exporters when Docker restarts

[dholt]

Summary

• Add PartOf=docker.service to the six containerized monitoring service units (node exporter, DCGM exporter, Slurm exporter, Prometheus, Grafana, Alertmanager).
• Root cause of node exporters don't work after initial run of slurm playbook #1267: the units declare Requires=docker.service without PartOf=, so an explicit Docker restart propagates as a one-way stop to the dependents and nothing starts them again. A first slurm-cluster.yml run on GPU nodes triggers exactly this — the NVIDIA container runtime configuration restarts Docker after the node exporter is already up — leaving exporters down until reboot.
• With PartOf=, a Docker restart propagates as a restart, so the units come back automatically (including after any manual systemctl restart docker).

Validation

• git diff --check; full role lint: 0 failures, 0 warnings.
• Live single-node GPU validation summary will be posted as a follow-up comment before review: fresh slurm-cluster.yml run, then asserting the node exporter is active at the end of the run and recovers after an explicit Docker restart.

Notes

• Fixes node exporters don't work after initial run of slurm playbook #1267 (stale-closed; the maintainer had twice marked it "not stale" — the failure mode is present on master until this change).
• One line per unit; no change to start ordering or normal operation.

[dholt]

Live validation summary

Validated on a disposable dual-GPU server (Ubuntu 22.04) with a fresh full slurm-cluster.yml deployment from a branch carrying this change:

• The rendered unit files contain PartOf=docker.service.
• Reproduction of the reported failure sequence: the node exporter was deployed and active, then the DCGM exporter playbook's NVIDIA container runtime configuration fired its restart docker handler — the exact sequence that leaves exporters dead on master. With this change the node exporter was active at the end of the run.
• Explicit restart recovery: systemctl restart docker on the node; both the node exporter and Slurm exporter units returned to active automatically, and :9100/metrics served node metrics.
• Slurm remained healthy throughout (node idle, srun --gpus=1 nvidia-smi passing).