feat: Edit LoRA section, text encoder / train_norm training, Graph an…

[DualChimerra]

…d sample previews

Edit LoRA (image-editing tools trained on before -> after examples, Cosmos reference-frame conditioning as in diffusion-pipe's Anima edit):

• runtime: reference latents concatenated on the temporal axis, loss on the target frame only; one or two input images; previews redo the first examples
• studio: project kind "edit" (migration v21), examples API (upload, text, bulk instruction, two tool types: image + text, image + image), phase gate, recommended config for edit versions and a reset to it, bundle export/import
• web: own sidebar section and tools list, examples page (grid + editor), overview and training page without fields edits don't use, reference images in Generate

Also included:

• text encoder LoRA and train_norm (q/k norm) training with ComfyUI export
• Graph board params manager and schema-driven params
• sample hover previews, Queue / Monitor sample feed changes
• Wiki section (pages/wiki)