fix(nodes): make per-node backend install async via gallery job queue

[localai-bot]

Summary

POST /api/nodes/:id/backends/install now returns HTTP 202 with a jobID immediately, instead of blocking the request for up to 3 minutes while the worker downloads and registers the backend. This unfreezes the React UI when installing on one or more nodes from the Backends picker.

The change is wired through the gallery service's existing async job queue, the same pattern /api/backends/install/:id already uses:

• galleryop.ManagementOp gains a TargetNodeID field so a single ManagementOp enqueued on BackendGalleryChannel can be scoped to one worker
• DistributedBackendManager.InstallBackend builds a one-element targetNodeIDs allowlist when TargetNodeID is set, reusing the same path UpgradeBackend already takes
• The HTTP handler enqueues the op, stores a node-scoped opcache row (node:<nodeID>:<backend>) so concurrent installs on different nodes don't collide, and returns { jobID, statusUrl, message }
• /api/operations now surfaces a nodeID field for node-scoped ops so the Operations panel can render attribution (and the bare backend slug shows in name instead of the prefixed key)
• NodeInstallPicker dispatches all installs in parallel, then polls /api/backends/job/:uid per job (1.5s interval, 6 min hard cap) until each settles; the modal stays closeable mid-install

Test plan

• In distributed mode, open the Backends picker on a meta backend, select one node, click Install: HTTP 202 with jobID (Network tab), row shows "Installing" immediately, Operations panel surfaces the job with nodeID, row eventually flips to "Installed"
• Install the same backend on two different nodes concurrently: two distinct jobs in the Operations panel, both complete independently, neither stomps the other's opcache row
• Install a backend whose gallery doesn't resolve on the target node: row flips to "Failed" with the worker's error in the tooltip; "Retry failed nodes" re-runs
• Verify backwards compatibility: POST /api/backends/install/:id (global, no node target) still fans out across the cluster unchanged
• Verify the picker modal can be closed while installs are in flight without losing visibility (Operations panel keeps tracking)

Follow-ups (not blockers)

• Orphan opcache rows are not auto-GCed if the handler's dispatch goroutine never drains (panic, process restart). /api/operations/:jobID/dismiss lets users clear them manually
• Concurrent duplicate dispatch of the same (nodeID, backend) will leak the first jobID in galleryService.statuses until process restart. JS picker dedupes via the selected Set so realistic UI flow is safe; no server-side dedupe added in this PR

Assisted-by: Claude:opus-4-7 [Edit] [Bash] [Agent]