Interpretable Neural Architecture Search via Bayesian Optimisation with Weisfeiler-Lehman Kernels