Add forward sampling methods that compute moments directly

[lucianopaz]

I've recently run into problems that have huge model variables. I can sample the posterior for the free random variables fine, but some deterministics and predictive variables are just to big to store their full posterior represented by chains and draws. We can develop a method that iterates through the free RV samples, draws or computes conditional values of variables of interest and instead of stacking them over chains and draws, compute their mean and std with single pass algorithms such as Welford's. We can even get histograms or quantiles if we were able to get a t-digest implementation that handles batch dimensions.