CARL: Causality-guided Architecture Representation Learning for an Interpretable Performance Predictor