Refactor: Portable characters & one canvas for training - #57

A character is a file now. Build one from reference images, train an adapter for it, and use the
same file on FLUX.2 and Krea 2.

One canvas. The Trainer tab is gone. Dataset, Caption and Train LoRA are nodes on the Studio
canvas, so one Run walks encode, train, attach and write in order. Existing projects migrate on
open (schema 21).

Character nodes. Encode Character, Compile References, Character to Dataset, Attach Adapter,
Write .char, Load Character, Edit Character. They replace the character side panel. A .char holds
one identity and a payload per model, so adding a model adds a payload instead of rebuilding the
character.

Models are visible. Nodes no longer resolve weights in the background. Character encoders,
training base checkpoints and Turbo de-distillation adapters all show in the missing-models popup,
which gained Download all, a queue and cancel. Exported graphs list every model with a download
link (recipe v2).

Canvas. Per-node run borders, search in the add-node menu, hover labels on port dots, and
minimum node heights from port count.

Models now live in Inline-Registry Repo