feat: add Video-MME-v2 benchmark task (#1289)

* feat(videomme_v2): add task config and default template

- Dataset: MME-Benchmarks/Video-MME-v2 (800 videos, 3200 questions)
- 8-option MCQ (A-H) with grouped non-linear scoring
- Generation config: max_new_tokens=64, temperature=0

* feat(videomme_v2): add scoring, prompts, and evaluation logic

- Grouped non-linear scoring: relevance (quadratic) + logic (chain-based)
- 3 group structures: [1,2,3,4], [1,[2,3],4], [[1,2],3,4]
- Answer extraction with 11 prefix patterns (A-H range)
- Per-level, per-category, per-group-type breakdown reporting
- Prompt aligned with official INSTRUCT_PROMPT
- Verified against VLMEvalKit implementation

* feat(videomme_v2): add subtitle variant (concatenated mode)

- Load word-level JSONL subtitles and prepend to prompt
- Graceful fallback when subtitle file is missing
- Task: videomme_v2_w_subtitle

* feat(videomme_v2): add reasoning mode variant

- Chain-of-thought prompt requiring Final Answer: <letter> format
- max_new_tokens=4096 for reasoning space
- Task: videomme_v2_reasoning

* fix(videomme_v2): report sub-category scores as separate metrics

Per review feedback (ref: PR #1285 pattern):
- Report relevance/logic group-type scores separately
- Report per-level (1/2/3) scores separately
- Refactor aggregate logic into _compute_all_subscores helper
- process_results returns same entry under all 6 metric keys
- Detailed second_head/third_head breakdowns still logged

---------

Co-authored-by: mwxely <mwxely@users.noreply.github.com>

Author: mwxely

lmms_eval/tasks/videomme_v2/_default_template_yaml

@@ -0,0 +1,46 @@
+dataset_path: MME-Benchmarks/Video-MME-v2
+dataset_kwargs:
+  token: True
+  cache_dir: videomme_v2
+  video: True
+test_split: test
+output_type: generate_until
+doc_to_visual: !function utils.videomme_v2_doc_to_visual
+doc_to_text: !function utils.videomme_v2_doc_to_text
+doc_to_target: "answer"
+generation_kwargs:
+  max_new_tokens: 64
+  temperature: 0
+  top_p: 1.0
+  num_beams: 1
+  do_sample: false
+process_results: !function utils.videomme_v2_process_results
+metric_list:
+  - metric: videomme_v2_score
+    aggregation: !function utils.videomme_v2_aggregate_results
+    higher_is_better: true
+  - metric: videomme_v2_relevance_score
+    aggregation: !function utils.videomme_v2_aggregate_relevance
+    higher_is_better: true
+  - metric: videomme_v2_logic_score
+    aggregation: !function utils.videomme_v2_aggregate_logic
+    higher_is_better: true
+  - metric: videomme_v2_level_1
+    aggregation: !function utils.videomme_v2_aggregate_level_1
+    higher_is_better: true
+  - metric: videomme_v2_level_2
+    aggregation: !function utils.videomme_v2_aggregate_level_2
+    higher_is_better: true
+  - metric: videomme_v2_level_3
+    aggregation: !function utils.videomme_v2_aggregate_level_3
+    higher_is_better: true
+lmms_eval_specific_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "\nAnswer with the option's letter from the given choices directly."
+  qwen3_vl:
+    format: "qwen3_vl"
+    pre_prompt: "Question: "
+    post_prompt: "Answer with the option letter only."
+metadata:
+  - version: 0.0