Scipy minimizer parameter covariance matrix calculation fails when parameter errors are too different

[JohannesGaessler]

If the errors of parameters are differ by ~10^4 the calculation of the Hessian matrix at the fit minimum with the SciPy optimize minimizer fails.
(The covariance matrix is calculated from the Hessian matrix.)
This is because Numdifftools seems to use only one step size for the calculation.
It might be necessary to implement the calculation of the hessian matrix ourselves or expose the step size to the user.
The described issue currently causes one of the examples to fail with the SciPy minimizer: #102 .