Implement numerical second derivative

[janosg]

Desired Situation

• We can calculate the second derivative of vector and scalar valued functions that take parameters as numpy arrays or DataFrames
• The three standard hessian formulae are supported. See ridout for formulae and statsmodels for a code example. We do not need complex steps.
• Function evaluations are done in parallel
• Interface and additional outputs similar to first_derivative
• Richardson extrapolation
• Check if it is possible to reuse evaluations returned by first_derivative

Proposed implementation

I suggest to do several independent PRs for this

• [x] Add basic functionality without richardson extrapolation building on the step generation functions we also use for first derivatives. It is important to design the code analogous to first_derivative, i.e. with parallel function evaluation and richardson extrapolation in mind. For now focus on numpy array interface.
• [x] Add pandas interface, dict outputs of func and error handling as in first_derivative
• [ ] Add richardson extrapolation
• [ ] Add intermediate outputs and visualization