allow adjustment of finite-difference epsilon for gradient checking?

[bbolker]

deriv12 (the function we use to compute finite-difference derivative & Hessian for gradient checking) has a default delta value of 1e-4. This may be what's messing up our Hessian calculations for large problems, e.g. see 3d set of figures here (mineigND is calculated with the numDeriv() package, i.e. using Richardson extrapolation, which is slow but much more accurate. Switch to Richardson extrapolation for that case, or at least allow an adjustment of the delta ... ?? (Or is this all scraping burned toast?)

I started to do this but got bogged down in the guts of lmer -> optimizeLmer -> optwrap -> deriv12 ... passing the control parameters through appropriately is getting unwieldy ...

[mmaechler]

still.. I think we should envisage to pass the numDeriv() control/method parameters. How much slower is the Richardson extrapolation? I "wildly" guess that using it may make the 'd'/'delta' less crucial for the achieved accuracy.