feat(omnidocbench): add normalized Levenshtein distance metric

[MaxwellJryao]

Add omnidocbench_nld_score metric computed as (1 - NLD) * 100, following the Kimi K2.5 technical report scoring method. The existing exact_match metric is preserved alongside the new one.

Summary

• Add omnidocbench_nld_score metric: (1 - normalized_levenshtein_distance) * 100, using the Levenshtein library
• When multiple reference answers exist, take the best (max) score across all answers
• Register the new metric in omnidocbench.yaml with aggregation: mean

In scope

• lmms_eval/tasks/omnidocbench/utils.py: add _normalized_levenshtein_score() helper, update omnidocbench_process_results to return both
  omnidocbench_exact_match and omnidocbench_nld_score
• lmms_eval/tasks/omnidocbench/omnidocbench.yaml: register omnidocbench_nld_score in metric_list

Out of scope

• No changes to other tasks (charxiv, ocrbench, ocrbench_v2, etc.)
• No changes to the existing omnidocbench_exact_match scoring logic
• No new dependencies added (Levenshtein is already declared in pyproject.toml under [project.optional-dependencies].metrics)

Validation

• python -m lmms_eval --model vllm --model_args model=Qwen/Qwen3-VL-8B-Instruct,tensor_parallel_size=2,data_parallel_size=4 --tasks omnidocbench
  --limit 4 --batch_size 4 | sample size: N=4 | key metrics: omnidocbench_exact_match, omnidocbench_nld_score both reported | result: pass

Risk / Compatibility

• Non-breaking: existing omnidocbench_exact_match metric is unchanged; the new metric is purely additive
• Results from prior runs remain valid and comparable

Type of Change

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