Fix confint ar1

[bbolker]

is there a reason to wait to merge this?

[bbolker]

OK, I see: Wald case is fixed but others aren't yet.

[kaskr]

We have to be careful not to apply this method too generally. Recall that it simply transforms the univariate confidence intervals, i.e. [a,b] => [f(a),f(b)].
This works fine for the ar1() parameter, because the internal parameterization is phi = f(theta). However, it wont work for more complex cases such as the correlation matrix elements of a us() term, because each such element is a function of multiple thetas.

[bbolker]

in the meanwhile there are some issues in here (in testthat/test-methods.R): there are both naming changes (which might be OK as long as we warn users in NEWS) but some further mismatches (the last row of the new version is bogus, should be the sigma CIs ...) Will have to dig in and see what's going on.

## new
> ci.2NB
                                         2.5 %     97.5 %
cond.(Intercept)                    5.48098714 5.58422551
cond.Days                           0.02481639 0.04289932
Subject.cond.Std.Dev.(Intercept)    0.06617725 0.15091785
Subject.cond.Std.Dev.Days           0.01134364 0.02635499
Subject.cond.Cor.Days.(Intercept)  -0.52088384 0.50221168
cond.Corr.Subject.Days:(Intercept) -0.52088384 0.50221168

## old/expected
> ci.2NB.expect
                                 2.5 %       97.5 %
cond.(Intercept)            5.48098714   5.58422551
cond.Days                   0.02481639   0.04289932
cond.Std.Dev.(Intercept)    0.06617725   0.15091785
cond.Std.Dev.Days           0.01134364   0.02635499
cond.Cor.Days.(Intercept)  -0.52088384   0.50221168
sigma                     183.81058474 444.73566864

[mebrooks]

I was just starting to try to resolve the conflicts here. Is that still worth doing for this branch?