chore(release): bump component versions for 26.07

[dholt]

Summary

• Bump component default versions for the 26.07 release train:
  • Slurm 25.11.6 → 26.05.1 (current upstream stable line from SchedMD)
  • GPU Operator chart v26.3.1 → v26.3.3 (default driver branch unchanged at 580.126.20)
  • Kubernetes GPU device plugin and GPU feature discovery charts 0.19.1 → 0.19.3
  • MIG manager (mig-parted) packages v0.14.1 → v0.14.2
  • Network Operator 26.1.1 → 26.4.0
  • Prometheus v3.11.3 → v3.13.0, Alertmanager v0.32.1 → v0.33.0, Grafana 13.0.1 → 13.1.0, kube-prometheus-stack chart 85.0.3 → 87.5.1
  • Spack v1.1.1 → v1.2.0
• Kubespray (v2.31.0), node exporter (v1.11.1), ingress-nginx chart (4.15.1), registry (3.1.1), and the MAAS role pin are already current and unchanged.
• Make the DEEPOPS_VERSION debug output in scripts/common.sh meaningful: default it to the checkout's git describe --tags --always with an unknown fallback, keep the config/env.sh override, and document the variable in config.example/env.sh.

Validation

• git diff --check origin/master..HEAD
• Full role lint: 0 failures, 0 warnings
• Ansible syntax-check suite across the standard playbook set
• Every bumped artifact verified against its primary source before the change: SchedMD tarball and mig-parted release packages return HTTP 200; Docker Hub manifests exist for the Prometheus, Alertmanager, and Grafana tags; the GPU Operator, device plugin, GFD, Network Operator, and kube-prometheus-stack chart versions are served by their Helm repositories.
• bash -n scripts/common.sh plus behavior check of the version derivation and env override.

Notes

• Draft until release QA gates complete: fresh Slurm deployment with a GPU job, Kubernetes/GPU Operator deployment with a CUDA pod, MAAS provisioning smoke, and upgrade-path validation. A sanitized QA summary will be attached before review.
• Slurm moves to a new upstream major line (26.05); the Slurm deployment gate below validates build and single-node GPU scheduling before this leaves draft.
• kube-prometheus-stack is a multi-major chart bump; monitoring deployment values are exercised during QA.

[dholt]

Live validation summary

All release QA gates for this PR completed on a disposable single-node NVIDIA H100 GPU server running Ubuntu 22.04:

• Slurm deployment (26.05.1): full playbooks/slurm-cluster.yml run finished with ok=304, changed=108, failed=0. Slurm built from the 26.05.1 source tarball; scontrol reports Version=26.05.1, the node joined the batch partition idle with gres/gpu=1, all Slurm services active, and the GPU job gate passed: srun --gpus=1 nvidia-smi saw the H100 on driver 580.159.03.
• Kubernetes / GPU Operator (v26.3.3, device plugin 0.19.3): playbooks/k8s-cluster.yml deployed cleanly; GPU Operator, device plugin, and node-feature-discovery pods running; the node advertised one allocatable nvidia.com/gpu; a CUDA smoke pod scheduled with a GPU request reached Succeeded and reported the H100 on driver 580.159.03.
• MAAS provisioning smoke: MAAS controller install passed (failed=0); services active; the MAAS API responded with version 3.5.12 and full capability list.
• Upgrade path: the direct 23.08-to-current in-place upgrade remains blocked by Kubespray's Calico minimum-version gate, as found and documented during the 26.05 release; the docs carry the explicit staged-upgrade/redeploy statement, which satisfies this gate.
• Static gates: full role lint (0 failures), Ansible syntax-check suite, OS compatibility and component audits, and public sanitization all passed; every bumped artifact was verified against its primary upstream source (tarball/package HTTP checks, container manifest checks, Helm repo index checks) before the change.

Not covered on this hardware: multi-node scheduling, Network Operator 26.4.0 install behavior (needs relevant NICs), and monitoring stack deployment values for the kube-prometheus-stack chart bump beyond syntax/render checks.