D-CoDe: Scaling Image-Pretrained VLMs to Video via Dynamic Compression and Question Decomposition