Supporting Sampling/Distribution-based Forecasting Evaluation

As implemented in #12, only quantile-based probabilistic forecasting is implemented. However, this thows away any possible information on correlations between time steps and/or variates (quantiles effectively model all of those as independent marginals). This was chosen since many TSFMs produce them anyways. In the future, ways to meaningfully evaluate sampling/distribution-based model outputs might better evaluate the quality of forecasts.