CounterAlign: Counterfactual Supervision for Vision-Language-Action Models