Add AGENTS.md operating guide and agent skills

[dholt]

Summary

• Add AGENTS.md at the repository root: an operating guide for AI agents (and new humans) with the repository map, first-time setup, Slurm and Kubernetes golden paths anchored on the validation tools from Add machine-readable deployment validation tools #1359, operating rules (validate don't assume, review inventories before running cluster playbooks, reboot awareness, check mode, idempotent reruns), and a list of known behaviors that look like failures but are not (Slurm login GPU hiding, Ansible fact caching, open vs proprietary kernel modules per GPU generation, uninitialized Kubespray submodule).
• Add skills/ with four self-contained procedures in the cross-tool agent-skills format (SKILL.md with name/description frontmatter): deploy-slurm-cluster, deploy-k8s-gpu-cluster, validate-gpu-cluster, and diagnose-driver-install. Each has preconditions, exact commands, expected outputs, and failure branches drawn from real GPU validation runs.
• Add a README pointer to AGENTS.md, the validation docs, and skills/.

Validation

• Doc-only change: git diff --check, YAML frontmatter parse for all four skills, public sanitizer pass on every added file.
• Inventory group names verified against config.example/inventory (kube_control_plane/etcd/kube_node, slurm-master/slurm-node).
• Every command in AGENTS.md and the skills was exercised during the Add machine-readable deployment validation tools #1359 live validation cycle on real GPU hardware (full Slurm 26.05.1 and Kubernetes/GPU Operator deployments plus the validation tools).

Notes

• The gotcha list and skill failure branches encode operational knowledge that previously lived outside the repository; this makes DeepOps operable end-to-end by an agent from a clean checkout.
• Follow-up candidates for the next train: a skill for airgapped deployments and one for MAAS-based provisioning.