Feat/ocrbench omnidocbench lang subset

[MaxwellJryao]

Summary

• Add separate Chinese/English subset score metrics for OCRBench_v2 (ocrbench_v2_accuracy_en, ocrbench_v2_accuracy_cn)
• Add separate Chinese/English/Mixed subset score metrics for OmniDocBench (omnidocbench_{exact_match,nld_score}_{en,cn,mixed})
• Fix spotting_evaluation crash when result["method"] is missing or not a dict

In scope

• New per-language aggregation functions and YAML metric entries for both benchmarks
• Refactor OCRBench_v2 bucket-filling logic into shared _fill_score_buckets helper
• Change OCRBench_v2 overall score from equal-weight EN/CN average to sample-count-weighted average
• Add _detect_document_language for OmniDocBench based on page_info.page_attribute.language (fallback: element-level majority vote)

Out of scope

• No changes to scoring logic (how individual samples are evaluated)
• No changes to dataset loading, prompts, or generation kwargs

Validation

Risk / Compatibility

• OCRBench_v2 overall score (ocrbench_v2_accuracy) now uses sample-count-weighted averaging instead of equal-weight EN/CN averaging; results will
  differ slightly on the full dataset (7400 EN vs 2600 CN)
• New metrics are additive; existing metric keys and values (except the weighting change above) are unchanged

Type of Change

• Bug fix (non-breaking change)
• New feature
• New benchmark/task
• New model integration
• Breaking change
• Documentation update
• Refactoring (no functional changes)