Bayesian R square for greta model

Inspired by the work from Gelman and others on [Bayesian R-square](https://stat.columbia.edu/~gelman/research/published/bayes_R2_v3.pdf), I was wondering if it would make sense to include such a function in greta or in an associated packages to run model checks on greta model fits.

In essence one needs two things:

- the vector of response values (y)
- the posterior draws from the linear predictor (which we could get from _calculate_) on the response scale (let's call this ypred)

Then in R code we have:

```r
# posterior residuals
e <- -1 * sweep(as.matrix(ypred), 2, y)
# variance in linear predictors
var_ypred <- apply(as.matrix(ypred), 1, var)
# variance in residuals
var_e <- apply(e, 1, var)
# R2 values
post_r2 <- var_ypred / (var_ypred + var_e)
```

The trick is basically to get the posterior draws from the linear predictors.

Any ideas?