CompoundPiece: Evaluating and Improving Decompounding Performance of Language Models