Krea2 generation & lora trainer

[imprsnst]

Generation

• Krea 2 Turbo and Krea 2 RAW canvas nodes, registered as Core builtins alongside Z-Image. Each declares requirements and a fit estimate so the model popup warns before a 26 GB load instead of failing mid-load.
• Reads ComfyUI-style files from Comfy-Org/Krea-2, bf16 builds only. The fp8_scaled, int8_convrot, mxfp8 and nvfp4 variants carry ComfyUI-specific scale tensors; the node says so up front. VAE is the diffusers-format one from Qwen/Qwen-Image. All repos public, no HF token needed.
• img2img support plus the usual take/progress plumbing.
• Extracted shared pipeline machinery from the Z-Image runner into models/pipeline_runtime.py and models/loaders.py. zimage/runner.py drops from ~1000 to ~250 lines and both families share one device/dtype/offload path. models/checkpoint.py centralises checkpoint inspection.

LoRA training

• Trainer is arch-aware (training/arch.py): z-image and krea2, each declaring base resolution, target modules and text-encoder handling.
• 4-bit (nf4) frozen base for Krea 2, which is what fits a 26 GB base into a 12 GB card at 512. The LoRA stays full precision, so the only cost is base fidelity. auto sizes base plus activations against the GPU and picks.
• New hyperparameters in the Trainer Adjust panel: base quantisation (auto/none/nf4), LoRA scope (full vs attention, per the Krea authors' advice for long runs), caption dropout, flip augmentation.
• raw base mode joins turbo_adapter/deturbo. Existing runs default to arch: z-image, so anything saved before this PR keeps working.
• Settings stage behind an Update button instead of applying as you type. Checkpoints encode the rank, targets and base they were built with, so applying a change to a resumable run asks first, then discards that run's checkpoints via the new training:discard channel. Finished LoRA files are never touched.

Project / licensing

• MIT to GPL-3.0-or-later (LICENSE, package.json, README badge).
• Added CODE_OF_CONDUCT.md, CONTRIBUTING.md, issue templates, PR template.
• README: Krea 2 section (files, VRAM floors, bf16-only and VAE-format gotchas), updated hardware table (Krea 2 1024² and LoRA training verified on an L40S 48 GB), updated known limits.