feat: add GEditBench benchmark with task type subsets

[davidberenstein1957]

Summary

Add GEditBench benchmark for image editing evaluation with 11 task type subsets.

Dataset

• Source: stepfun-ai/GEdit-Bench (HuggingFace)
• Task types: background_change, color_alter, material_alter, motion_change, ps_human, style_change, subject_add, subject_remove, subject_replace, text_change, tone_transfer
• Collate: prompt_with_auxiliaries_collate

Implementation

• Add setup_gedit_dataset in src/pruna/data/datasets/prompt.py
• Support subset param for filtering task types
• Filter to English instructions only
• Register in base_datasets
• Add BenchmarkInfo entry with metrics: ["accuracy"], subsets list
• Auxiliaries should include image (input_image), subset
• Add test

Acceptance

• PrunaDataModule.from_string("GEditBench") works (all subsets)
• PrunaDataModule.from_string("GEditBench", subset="background_change") works
• Auxiliaries include image, subset fields
• Test passes