Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models