What Constitutes Good Contrastive Learning in Time-Series Forecasting?