Allow use of robust versions of priors when prior is 'saturated'

[johnrbryant]

A prior is 'saturated' if the associated main effect or interaction has the same length as data 'y'. At present only priors with normally-distributed errors are permitted. To allow t-distributed errors, we need to revisit the updating of 'sigma'.