On the High-dimensional Power of Linear-time Kernel Two-Sample Testing under Mean-difference Alternatives