REGEN: Learning Compact Video Embedding with (Re-)Generative Decoder