Slight improvement n.init

[BertvanderVeen]

Apologies, but the improvement for n.init required a bit further fine tuning.

• n.init.max should of course be < n.init (so, by default now at 10).
• slightly more informative warning when n.init.max is reached
• if a model would be accepted by the norm of the gradient, first check if the two gradient differ sufficiently, otherwise there is no point. This gives a slightly more strict convergence criteria for n.init.max - it would be possible to continue accepting models indefinitely otherwise.

I also uncommented the if(TMB)trace<-FALSE in the gllvm.R file, since this tracing can actually be useful in combination with n.init.