How Much Temporal Long-Term Context is Needed for Action Segmentation?