Comparing Normalizing Flows with Kernel Density Estimation in Estimating Risk of Automated Driving Systems