CounterVQA: Evaluating and Improving Counterfactual Reasoning in Vision-Language Models for Video Understanding