Allow "R-side" covariance structures

[jake-westfall]

It would be nice to be able to support general covariance matrices (or, you know, at least a small number of popular presets) for the errors in Gaussian models -- commonly called the R matrix in the mixed model literature (to be distinguished from G, the cov matrix of the random effects). The most obvious structure that we'd want to accommodate would be serial autocorrelation, for example AR(p) structures within the levels of some factor.

Similar functionality for non-Gaussian responses would, I think, be quite tricky. I'd be curious to see what other packages support.

[Gattocrucco]

I sometimes would like at least to have a different error variance parameter on different groups.