Token-Count-Based Batching: Faster, Cheaper Embedding Inference for Queries