Leveraging Text-to-Video Diffusion Models for Synthetic Gait Dataset Generation and Appearance-Robust Person Recognition