Close install-review gaps: vision model auto-pull, GHCR visibility, manual steps

[koydas]

• Add qwen2.5vl:3b to apps/ollama/application.yaml's auto-pull list so the
  vision model ollama-chat routes image requests to (src/lib/conversations.js
  VISION_MODEL) is fetched on sync like the text models already are.
• Document GHCR package visibility for the three git-source Applications
  (ollama-chat, homelab-gateway, piper): confirmed all three currently
  public via anonymous tags/list pulls, none have an imagePullSecret, and a
  fresh package defaults to private regardless of the source repo's
  visibility. Add ADR-0018 recording that as a deliberate decision (mirrors
  the example already called out in ollama-chat's own ADR-0006, but that
  was never carried over here or to homelab-gateway).
• Add a "Manual post-install steps" checklist consolidating the MetalLB
  DHCP reservation, Grafana admin Secret, kube-prometheus-stack CRDs, GPU
  time-slicing ClusterPolicy patch, and GHCR visibility check with exact
  commands in one place.
• Fill in README's Structure section for apps/ollama-chat, apps/whisper,
  apps/piper, apps/homelab-gateway, and apps/gpu-time-slicing, which were
  missing even though docs/architecture.md already documented them —
  verified piper's Application stays consistent with piper-tts-server now
  that it's documented as its own separate repo.