gpu — docs: usage/gpu.md + setup/gpu.md + architecture/gpu-visibility.md + Helm values reference

[gnana997]

Part of #202.

Context

Operator-facing docs for the v1.2 GPU surface, plus the Helm values reference for periscope-gpu-agent.

Acceptance

• docs/usage/gpu.md — Pod↔GPU map walkthrough + Idle GPU finder + DCGM reconciler. Screenshots from the v1.2 capture session. Same structure as docs/usage/cve.md.
• docs/setup/gpu.md — prerequisites (GPU Operator install, DCGM exporter, optional DRA driver), Helm values for periscope-gpu-agent, IAM block needed for cluster-wide instance pricing reads (none — pricing is static), RBAC requirements
• docs/architecture/gpu-visibility.md — public version of the design memo (adapted from the private PRD). Covers the spine architecture, the four data sources, the per-node DaemonSet rationale, the DCGM reconciliation strategy.
• docs/setup/cluster-rbac.md — new section "GPU visibility (optional, v1.2+)" with the K8s RBAC + IAM additions
• Sync-docs script updates: scripts/sync-docs.mjs needs to know about the new doc paths
• Update README.md features section with a GPU bullet

Files this likely touches

• New: docs/usage/gpu.md, docs/setup/gpu.md, docs/architecture/gpu-visibility.md
• Update: docs/setup/cluster-rbac.md, docs/usage/meta.json, docs/setup/meta.json, docs/architecture/meta.json
• Update: README.md
• Update: scripts/sync-docs.mjs (if any path-rewriting is needed)

Dependencies

Land after the user-facing sub-tasks are mostly stable so screenshots don't churn.

Notes for contributors

• Match the voice + structure of docs/usage/cve.md (Inspector v2 doc) — same operator-question shape, same screenshot rhythm
• The architecture doc is the public version of the private PRD; trim the strategic / hosted.ai bits, keep the technical sections
• Helm chart reference doc goes in docs/setup/values.md (existing file)