In the process of optimizing Spark performance for a machine learning project, which technique is used to precompute and store intermediate results, thereby reducing computational overhead in subsequent tasks?