SimTS: Rethinking Contrastive Representation Learning for Time Series Forecasting