fix: improve evaluation logic across 10+ existing benchmarks

[Luodian]

Summary

Fixes and enhancements across multiple existing benchmark tasks to improve evaluation accuracy and code quality.

Changes by benchmark

Task	Change
egotempo	Enhanced temporal reasoning evaluation with detailed per-category metrics
lvbench	Simplified evaluation logic, removed redundant option-matching code
mlvu	Streamlined answer extraction, removed unused imports
mmstar	Simplified MCQ extraction with cleaner regex patterns
omnidocbench	Major refactor — improved document evaluation metrics, better text/table/formula handling
realworldqa	Cleaned up YAML config (removed hardcoded options), improved utils
spatialviz	Added missing metric aggregation functions
videomme	Simplified video evaluation, removed redundant frame extraction code
videommmu	Enhanced video processing with better multi-image support
mmlongbench_doc	Improved document benchmark with better page-level evaluation

Test plan

• Run affected benchmarks with a compatible model to verify no regressions
• Spot-check omnidocbench refactor (largest change) with document-capable model
• Verify realworldqa yaml changes don't break task loading