Watch for boundary cases from fast=true in glmm fits

[dmbates]

If one of the constrained elements of θ converges to zero in the fast=true stage when fitting a glmm then it should be given a nonzero starting value in the general optimization.

[dmbates]

I suggested on a private slack channel that it will be easier to simply remove the two-stage optimization. That is, when fast=true is set the optimization proceeds as done currently but if fast=false, the default, then the starting estimates are simply the fixed-effects from the GLM fit and the value of theta that corresponds to an identity matrix.

I created the two-stage optimization because I am usually paranoid about obtaining good starting estimates in an optimization problem with more than a few unknowns. However, it doesn't seem to be all that important now as the BOBYQA algorithm implemented in the NLopt package is very reliable on these problems.