feat: add VBench video generation evaluation benchmark

[Luodian]

Summary

• Add VBench and VBench 2.0 benchmark suites for comprehensive video generation evaluation
• Covers 40+ task variants: subject/background consistency, motion quality, spatial relationships, aesthetic quality, human action/anatomy preservation, dynamic attributes, material/mechanics/thermotics, and more
• Includes build_dataset.py for dataset preparation from VBench prompts

References

• VBench: https://arxiv.org/abs/2311.17982
• VBench 2.0: https://arxiv.org/abs/2503.21755

Task list

• vbench (group): VBench 1.0 tasks (15 variants)
• vbench2 (group): VBench 2.0 tasks (13 variants)
• vbench_all: Combined VBench 1.0 + 2.0

Test plan

• Verify task registration with lmms-eval --tasks list | grep vbench
• Run a single VBench task with a video generation model
• Confirm scoring metrics align with VBench reference implementation