fix: harden GPU runtime convergence

[dholt]

Problem

Several GPU-cluster paths diverged as newer DGX OS, Slurm, and Ubuntu releases
changed their service, module, package, and security defaults. In combination,
these gaps prevent a complete Slurm deployment with monitoring and Pyxis from
converging and validating reliably.

Changes

• Select the correct SSH service when applying the Slurm login GPU guard.
• Transition DGX systems safely from legacy nv_peer_mem to the in-tree
  nvidia_peermem module.
• Make exclusive-job detection custom-prefix aware and fail closed when Slurm
  queries fail or return malformed data.
• Install a command-scoped AppArmor profile for Enroot instead of changing the
  host-wide unprivileged-user-namespace policy.
• Select and verify the DGX OS 7 DCGM 4 package on Ubuntu 24.04 and newer.
• Put custom-prefix Slurm client binaries on the exporter container's PATH as
  read-only mounts.
• Avoid repeating a direct-login driver test after Slurm has intentionally
  hidden GPUs from SSH; allocated srun validation remains authoritative.

This incorporates and hardens the useful changes proposed in #1389 through
#1395. The GPU power/clock batching from #1393 remains separate because it is
an optimization rather than part of this correctness fix.

Validation

• python3 -m unittest discover scripts/validation/tests — 25 tests passed,
  including failed, empty, malformed, exclusive, and last-user Slurm query
  cases.
• Full Ansible role lint passed with ansible-lint 26.1.1.
• Syntax checks passed for the Slurm, Kubernetes, monitoring, GPU, and MAAS
  playbooks.
• Public-content sanitization and the complete non-live release QA pipeline
  passed.
• Exact commit bc571ff9 passed a fresh Ubuntu 24.04 GPU-backed Slurm
  deployment and its immediate convergence rerun. The final playbook completed
  with failed=0; the validator proved direct-login GPU isolation, allocated
  srun GPU access, a populated custom-prefix Slurm exporter scrape with zero
  collector errors, a GPU container through Pyxis/Enroot, and correct
  exclusive/nonexclusive prolog selection.
• The live run also loaded the command-scoped AppArmor profile while preserving
  the host-wide unprivileged-user-namespace restriction.

All 30 public CI checks pass at the validated commit.