krea2 with qwen embed based reference image training and opt-in clean ref latent training

[bghira]

This pull request adds comprehensive support for the Krea 2 model family, including LoRA fine-tuning, configuration, and integration into the model registry and UI. It introduces new model components, configuration options, and tests to ensure correct behavior. The main changes are grouped below:

Krea 2 Model Integration:

• Added Krea 2 model support to the model registry and metadata, including class definitions, module paths, and available flavours (raw, turbo). (simpletuner/helpers/models/model_metadata.json, simpletuner/helpers/models/common.py) [1] [2] [3]
• Introduced new Krea 2 model components: Krea2, Krea2Pipeline, Krea2Transformer2DModel, and Krea2LoraLoaderMixin, making them available for import. (simpletuner/helpers/models/krea2/__init__.py, simpletuner/helpers/models/krea2/lora_pipeline.py) [1] [2]
• Added the Krea2PipelineOutput dataclass for pipeline outputs. (simpletuner/helpers/models/krea2/pipeline_output.py)

Configuration and UI:

• Added a new configuration field krea2_reference_latents to enable reference-dataset training, with UI integration and documentation. (simpletuner/simpletuner_sdk/server/services/field_registry/sections/logging_fields.py)
• Provided an example configuration for Krea 2 LoRA training. (simpletuner/examples/krea2.peft-lora/config.json)

Testing and Validation:

• Added a dedicated test suite for Krea 2, covering model registry resolution, configuration parsing, LoRA loader behavior, pipeline call signatures, and reference latents handling. (tests/test_krea2_model.py)