TATA: Stance Detection via Topic-Agnostic and Topic-Aware Embeddings