At-the-Roofline Sparse Tensor Contractions on Vector Processors for Inference