fix: make memory fact locale-safe and integer-valued

[dholt]

Summary

• Make the memory custom fact locale-safe and integer-valued: compute total_mb with awk printf "%d" instead of the default float print.
• Root cause of AnsibleUndefinedVariable: ''ansible.utils.unsafe_proxy.AnsibleUnsafeText object'' has no attribute ''total_mb''' #1321: awk's default output renders floats using the host's LC_NUMERIC decimal separator. On comma-decimal locales the fact emitted { "total_mb": 60760,1 } — invalid JSON — so Ansible silently provided the fact as a raw string and slurm.conf templating failed with the reported AnsibleUnsafeText ... has no attribute 'total_mb' error. This also explains why the error only reproduced for some users: it depends on the host locale.
• Harden the slurm.conf template to parse a string-typed fact (from_json) so any future malformed fact fails clearly instead of with an opaque attribute error.

Validation

• bash -n on the fact script; the fact output parses as JSON with an integer total_mb.
• Demonstrated the defect mechanism: awk '{print $2*0.95}' emits 60760.1 (locale-dependent separator), printf "%d" cannot emit a separator.
• ansible-playbook --syntax-check playbooks/slurm-cluster.yml; full role lint: 0 failures.
• Live single-node GPU validation summary will be posted as a follow-up comment before review: fresh Slurm deployment asserting RealMemory templates as an integer and the fact parses under a comma-decimal locale on the deployed node.

Notes

• Fixes AnsibleUndefinedVariable: ''ansible.utils.unsafe_proxy.AnsibleUnsafeText object'' has no attribute ''total_mb''' #1321. printf "%d" truncates toward zero, preserving the intended 5% headroom; output is unchanged on hosts where the value was already integral.

[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 deployed /etc/slurm/slurm.conf templated RealMemory=244707 — an integer computed by the new fact.
• On the deployed node, a comma-decimal locale (de_DE.UTF-8) was generated and the fact executed under it: output parsed as valid JSON with an integer total_mb ({ "total_mb": 244707 }). Under the same conditions the previous implementation emits a comma-decimal float, which is invalid JSON and reproduces the reported failure.
• Slurm deployed and scheduled normally (srun --gpus=1 nvidia-smi passing).