Self-Distilled Representation Learning for Time Series