make computations around invhess more robust numerically.

for example removes many occurrences of -Inf in computing Laplace approximations by scaling invhess for computing log of determinant.