Add fine-tuning support for LightOnOCR-1B-1025 model

Description

Describe the feature
Please add support for fine-tuning the [LightOnOCR-1B-1025](https://huggingface.co/lightonai/LightOnOCR-1B-1025) model.
This model is an end-to-end differentiable vision-language OCR model developed by LightOn AI, capable of performing OCR, layout analysis, and multilingual document understanding.
The model card mentions upcoming Transformers integration and support for LoRA/domain fine-tuning, but currently there’s no standard fine-tuning interface or example script available.

Fine-tuning support (either native or through transformers + peft integration) would enable developers to adapt this model for specific document domains — receipts, forms, scientific articles, invoices, etc.

Paste any useful information

• Model card: https://huggingface.co/lightonai/LightOnOCR-1B-1025

• Blog reference: https://huggingface.co/blog/lightonai/lightonocr

• Suggested integration targets:

  • transformers.VisionEncoderDecoderModel subclass
  • PEFT/LoRA adapter configuration for efficient fine-tuning

Additional context
Fine-tuning support will allow users to:

• Adapt OCR for domain-specific layouts or languages
• Experiment with low-resource adaptation via LoRA
• Easily share fine-tuned checkpoints on Hugging Face
• Increase adoption of LightOnOCR across document AI research and production use-cases

If possible, please also include an example notebook or script demonstrating minimal fine-tuning workflow using the ModelScope or Hugging Face ecosystem.

Start by reading the LightOnOCR-1B-1025 model card and blog reference, then evaluate the suggested VisionEncoderDecoderModel integration and PEFT/LoRA adapter configuration. Done should mean a standard fine-tuning interface plus a minimal ModelScope or Hugging Face notebook or script demonstrating the workflow.