What is the computational complexity of standard self-attention for a sequence of n tokens with dimension d?