Foldable SuperNets: Scalable Merging of Transformers with Different Initializations and Tasks