Implement (with a warning) sensitivity methods using finite differences

[ben18785]

For sensitivities, this requires 2x number of parameters per iteration if we use central differences; 1x the number of parameters if we use forward/backward differences.

For 2nd order sensitivities, more function evaluations are required.

[ben18785]

This is a good reference on finite differences.

Note that I don't expect this approach to be as computationally efficient as using automatic differentiation or calculating the sensitivities outside of Pints. It will, however, allow users to quantify, in terms of the number of effective samples per iteration not computational time whether it is worth solving for sensitivities.

[MichaelClerx]

I have some serious doubts about this!

Apart from my gut feeling (backed up by some bad experiences) that they're a terrible idea for serious problems, there's the question of where do we stop with Pints.
Yes it's something we could add, but I'm much happier saying "pints is for optimising / sampling" and leaving the rest for the users and other specialist packages. (See #501 )

[MichaelClerx]

This is related to #101

[MichaelClerx]

Example of how this could be implemented in a wrapper: #892