Theoretical Constraints on the Expressive Power of $\mathsf{RoPE}$-based Tensor Attention Transformers