gpu — frontend: Pod ↔ GPU map view (cluster → node → GPU → MIG/slice → bound pods + filters + sparklines)

[gnana997]

Part of #202.

Context

The headline operator view. A fleet-wide table showing every GPU-bound pod across every cluster Periscope manages, with the topology hierarchy (Cluster → Node → PhysicalGPU → MIGSlice or TimeSlice → BoundPod). Click a row → live util sparkline + neighbor pods on the same device + K8s events + logs link.

Acceptance

• New route at /clusters/<c>/gpu and a top-level /gpu for the fleet-wide view
• Sidebar entry under "EKS" group, auto-detected by whether any node carries nvidia.com/gpu.product labels
• Table: cluster · node · GPU index · model · MIG profile / time-slice / DRA · bound pod(s) · namespace · owner labels · current util
• Filters: cluster, GPU model (A100/H100/L4), MIG profile, namespace, util threshold
• Click row → detail pane with: live util sparkline (last 5min), neighbor pods on same device (for MIG / time-slice), K8s events, link to logs
• Live SSE updates (allocation changes + metric refreshes)
• Empty state when no GPU nodes detected — explain prerequisites + link to setup docs
• No new audit verb — reads are not audited (CloudTrail covers IMDS reads if enabled)

Files this likely touches

• New: web/src/pages/GPUMapPage.tsx
• New: web/src/components/gpu/GPUMapTable.tsx, GPUDetailPane.tsx
• New: web/src/hooks/useGPUMap.ts
• New backend handler: /api/clusters/<c>/gpu/map

Dependencies

Blocks on:

• gpu — backend: per-node DaemonSet (Epic: GPU + AI workload visibility — Pod↔GPU map, Idle GPU finder, DCGM reconciler (v1.2) #202 sub-task above) — needs allocation snapshots
• gpu — backend: DCGM scraper (Epic: GPU + AI workload visibility — Pod↔GPU map, Idle GPU finder, DCGM reconciler (v1.2) #202 sub-task above) — needs util / FB / power metrics

Notes for contributors

• Mirror the existing Karpenter dashboard pattern — same auto-detect-by-CRD-or-label, same resizable detail pane
• Sparkline component already exists in the SPA for the Karpenter cost view; reuse
• Time-slicing case: N pods share one device → render as a single row with all bound pods listed + per-pod util when DCGM_FI_PROF_GR_ENGINE_ACTIVE is available; otherwise "shared, util N% across M pods" honestly
• MIG case: render each slice as a sub-row of the physical GPU