Universal Assisted Generation: Faster Decoding with Any Assistant Model