v1.2.68 - Run queue and activity centre, models browser & mid-run LoRA snapshots

v1.2.68

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

Projects opened in this version are upgraded to schema 18 and will not open in
older versions.

Run queue and activity centre

• New Active Runs control in the header: live count, queue positions, per-run
  cancel, cancel all, and recent history.
• Runs keep going across a tab switch and a project switch, and the UI re-syncs
  after a reconnect.
• A connection dot shows whether Core is reachable.
• Runs submitted straight to the Core API appear in the same place, and there is
  a new GET /v1/runs endpoint.

Models

• New Models tab in the sidebar: a read-only tree of every models root, with
  sizes.
• --models-dir now adds to ./models instead of hiding it.
• Files added or removed on disk reach the model pickers without restarting
  Core.

Graph and canvas

• Run control gains a dropdown: Copy graph JSON, Export JSON, Duplicate graph.
• An exported graph can be dropped back onto the canvas to rebuild it.
• Nodes whose picked model or LoRA is missing are marked with a red border.

Training

• Mid-run LoRA snapshots are listed per run, with download and Add to models so
  you can test one without waiting for the run to finish.
• The encoded dataset is cached on disk, so resuming does not re-encode it.
• A usable LoRA is kept at every checkpoint and whenever a run is stopped.

Fixes

• A run's output could land in whichever project was open when it finished
  rather than the one it started in.
• Re-running a node made it uncancellable.
• A cancelled run left its node spinning forever.
• MiniMax H3 ignored cancel for tens of seconds.
• Cancel no longer claims a run has stopped before it actually has.
• Restarting Core now clears runs it no longer knows about.
• Duplicating a graph copies the wiring, and copy/paste works for generation
  nodes.
• The text encoder is no longer moved to the CPU on cards with plenty of VRAM,
  which was costing two multi-GB copies per render.

Upgrading

git fetch && git pull
./webui.sh --install --extra all     # Windows: .\webui.bat --install --extra all
./webui.sh                           # Windows: .\webui.bat

What's Changed

• Run queue and activity centre, models panel, training snapshots by @imprsnst in #45

Full Changelog: v1.2.67...v1.2.68