- Use (old default) mode_compute=4 which is closer to the algorithm
   used by Frank Schorfheide and ensures that the hessian matrix is well
   behaved (contrary to the new default, because of the asymptote at 0
   in the beta prior for autoregressive parameter ρ).

 - Change parameterization for mst. A normal prior on mst is not
   equivalent to a normal prior on log(mst) (which is done the
   parameterization in the JAE paper).

Closes #2177.