In the context of optimizing Spark performance for a machine learning project, which technique is used to combine multiple small tasks into larger tasks to reduce scheduling overhead and enhance processing efficiency?