feat: deploy/find-capacity/benchmark/troubleshoot agent skills

[Dexhub]

What

Four agent skills for deploying open models on EKS, mirrored across .claude/skills/ (Claude Code), .kiro/skills/ (Kiro), and AGENTS.md (generic agents):

• deploy-model — resolve a registry entry, check capacity, ensure the hf-token secret, generate a blueprint, apply it, and verify the /v1/chat/completions endpoint.
• find-capacity — map a GPU request to an instance type and run the capacity scanner.
• benchmark-model — concurrency sweep against a deployed endpoint (tokens/sec, TTFT, ITL, cost).
• troubleshoot-inference — diagnose OOM, CUDA mismatch, Karpenter scaling, and gated-repo/HF-token failures.

They drive the registry/, aoe-capacity, and aoe-blueprint capabilities via natural-language requests.

Scope / safety

Additive only; updates root README. No source behavior changes. Skills are documentation/instruction files — no executable code. Facts are verified-or-unset: the skills instruct agents never to invent benchmark numbers or a tool_parser.

Note

The .kiro/skills/<name>/SKILL.md layout matches the repo's existing Kiro skills (nodepools, argocd-addon). These skills reference aoe-capacity / aoe-blueprint, which land in their own PRs.

[Dexhub]

@fhsilva Review requested: this PR is part of the Open Model Factory series. You now have read access to the Dexhub/ai-on-eks fork; please follow and review this slice here.