Needed feature: Finite-difference testing of Cartesian gradient

[leeping]

geomeTRIC will fail to converge if the underlying code is computing the incorrect gradient. The failure mode is that the optimizer takes very small, bad-quality steps.

When this happens, geomeTRIC should trigger a finite-difference test of the Cartesian gradient. This will have the benefit of aiding troubleshooting, i.e. if the Cartesian gradient is wrong, then it shows the underlying gradient code as the source of the problem, and not geomeTRIC itself.