Krea 2 LoRA DreamBooth trainer

[apolinario]

What does this do?

Adds LoRA support for Krea 2 (now merged in #14045):

• examples/dreambooth/train_dreambooth_lora_krea2.py — DreamBooth LoRA trainer with the authors' recommended defaults (rank/alpha 32, target layers, lr 3e-4 constant), fp8 support, validation, plus README_krea2.md and requirements_krea2.txt.
• Krea2LoraLoaderMixin in loaders/lora_pipeline.py; Krea2Pipeline now inherits it so LoRA adapters can be saved/loaded.
• Krea2Transformer2DModel.forward and the pipeline __call__ accept attention_kwargs for runtime LoRA scaling.
• LoRA loader tests in tests/lora/test_lora_layers_krea2.py.

Ported from the Krea 2 staging repo and rebased onto current main.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[linoytsaban]

[linoytsaban]

Note on the latest commit: after porting the pipeline to main, training crashed because the trainer calls Krea2Pipeline._pack_latents(...) and Krea2Pipeline._unpack_latents(...) at the class level, and these are now instance methods (they use self.patch_size / self.vae_scale_factor) - for context they were @staticmethod before the merge iirc. So the first positional argument bound to self, shifting the rest - _pack_latents raised TypeError

Fix: I inlined the two small pack/unpack ops directly in the training loop (patch_size=2), to work around it.

Alternative (maybe in a separate PR): restoring @staticmethod on _pack_latents/_unpack_latents (as in QwenImagePipeline, which they're # Copied from).

[DN6]

@bot /style

[github-actions]

Style fix runs successfully without any file modified.