Large-Scale Actionless Video Pre-Training via Discrete Diffusion for Efficient Policy Learning