Cost-efficient Gaussian tensor network embeddings for tensor-structured inputs