Can Video Large Multimodal Models Think Like Doubters-or Double-Down: A Study on Defeasible Video Entailment