ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting