DADiff: Diffusion-Driven Cross-Domain Policy Adaptation for Reinforcement Learning