Guided by Stars: Interpretable Concept Learning Over Time Series via Temporal Logic Semantics