[features] Input QA prompt to VLM

[LukeLIN-web]

https://research.nvidia.com/labs/dir/pbench/

Support such features:
generate video, then

ask VLM

Question: Is the large truck positioned directly in front of the vehicle throughout the video?
Category: Space: Relationship
Answer: yes
Correct Answer: yes
Score: 1

Question: Does the truck veer off the road before coming to a stop on the shoulder?
Category: Time: Order
Answer: no
Correct Answer: yes
Score: 0

It is useful to evaluate some tasks such as #44

[LukeLIN-web]

we could support the datasets which don't have final image but have answer such as https://cbrower.dev/vpct in this way.

[LukeLIN-web]

TODO:

• change gpt4o evaluation
• draw a figure in readme to illustrate.
• update readme.md