feat: add COVER and WM-aBench video understanding benchmarks

[Luodian]

Summary

• Add COVER (Counterfactual Video Reasoning, ACL Findings 2025) benchmark for causal video understanding
• Add WM-aBench (World Models aBench) with 36+ task variants for comprehensive world model evaluation

Benchmarks

COVER

Tests causal understanding in videos via counterfactual question generation. Includes generate_qa.py for automatic QA pair generation from video annotations.

WM-aBench

Evaluates world model capabilities across multiple dimensions:

• Spatial: relative position, occupancy, multiview
• Motion: direction, speed, trajectory
• Physical: mechanistic knowledge, compositionality
• Temporal: extension, positioning, transitivity
• Visual: attribute recognition (color, shape, material)
• Counting: discrete counting, relative counting

Uses ManISkill, TDW, Physion, Habitat, and CARLA simulation environments.

Test plan

• Verify task registration with lmms-eval --tasks list | grep -E "cover|wm_abench"
• Run COVER with a video-capable model
• Run a WM-aBench subset task
• Confirm group yaml correctly aggregates all subtasks