Fast and Memory-Efficient Subspace Embeddings for Tensor Data with Applications