Semantic Token Reweighting for Interpretable and Controllable Text Embeddings in CLIP