At the moment, the only way to work with vectors is multivariate distributions.
Thanks to Distributions.product_distribution, we can also combine univariate distributions. However, as shown in #14, it is rather inefficient. Maybe we want to allow univariate distributions with a size argument for sampling.