[BENCHMARK] Add DPG (Descriptive Prompt Generation) benchmarks

[davidberenstein1957]

Description

Add DPG (Descriptive Prompt Generation) benchmarks for image understanding evaluation.

Details

• Source: GitHub CSV from TencentQQGYLab/ELLA repo
• Subcategories: entity, attribute, relation, global, other
• Collate: prompt_with_auxiliaries_collate

Implementation

• Add setup_dpg_dataset in src/pruna/data/datasets/prompt.py
• Support category param for filtering subcategories
• Register in base_datasets
• Add BenchmarkInfo entry with metrics: ["qa_accuracy"], subsets list
• Auxiliaries should include questions list and category_broad for evaluation
• Add test

Acceptance Criteria

• PrunaDataModule.from_string("DPG") works (all subcategories)
• PrunaDataModule.from_string("DPG", category="entity") works
• Auxiliaries include questions and category_broad fields
• Test passes