Run queue and activity centre, models panel, training snapshots

[imprsnst]

Adds a real notion of "a run" to the UI, a read-only models browser, graph
import/export, and surfaces the LoRAs training already writes mid-run.

Schema bumps to 18 (new generation_runs table). Version bumped to 1.2.68.

Run queue and activity centre

• New "Active Runs" control in the workspace header: live count, queue
  positions, per-run cancel, cancel all, and recent history.
• The queue lives on RunManager behind an observer seam, so runs submitted
  straight to POST /v1/runs show up and cancel from the same place. New
  GET /v1/runs endpoint.
• Runs now survive a tab switch and a project switch. Event subscriptions moved
  up to App.tsx, and the client re-syncs on reconnect.
• Connection dot showing whether Core is reachable, with Disconnected and
  Reconnected notices.

Fixes found along the way

• A run's output landed in whichever project was open when it finished, not the
  one it was started in. The project is now pinned at submit.
• Re-running a node deleted its own tracking, leaving it invisible and
  uncancellable.
• A cancelled run broadcast nothing, so the node span forever.
• MiniMax H3 ignored cancel for tens of seconds while it paid ~40GB of restore
  transfers on the way out.
• Cancel no longer claims a run stopped before it has. Nodes say "Stopping..."
  until Core confirms.
• A Core restart now clears runs it no longer knows about.
• The text encoder was parked on the CPU even on cards with plenty of VRAM,
  costing two multi-GB copies per render for nothing.

Models

• New Models tab in the canvas sidebar: read-only tree of every models root,
  with sizes. No file actions.
• --models-dir is additive, so a custom root no longer hides ./models.
• Model files added or removed on disk now reach the pickers without restarting
  Core, via a refresh beside every model select and on window focus.

Graph and canvas

• Run control gains a dropdown: Copy graph JSON, Export JSON, Duplicate graph.
• Exported graphs can be dropped back onto the canvas to rebuild.
• Duplicate graph now copies the wiring, and copy/paste works for generation
  nodes at all (it silently did nothing before).
• Nodes whose picked model or LoRA is not installed get a red border.

Training

• Mid-run LoRA snapshots are listed per run, with download and "Add to models"
  so they can be used in a graph.
• Encoded dataset is cached on disk, so a resume does not re-encode.
• Checkpoint setting is a checkbox, and a usable LoRA is kept at every
  checkpoint and on stop.