log crossprodexp

[rmcminds]

implement a way to normalize the variance multipliers in numerically stable way, and save the normalized result rather than just passing it on. the normalized result should be identified but the raw one isn't, and that could be what's leading to terrible nEff for those results.

[rmcminds]

might be possible while you're at it to implement the equivalent of a sum-to-zero constraint on those parameters, too, rather than the currently used soft centering plus forced normalization. the soft centering is also a likely candidate for bad mixing and the performance of the whole model might improve drastically...