Add support for autodiff

[nilsvu]

Automatic differentiation would be useful so we don't have to implement linearizations or Jacobians of complicated analytic functions. Examples where autodiff would help:

• Jacobians and Hessians of coordinate maps. As a first step we can use autodiff just to test our implementations of map Jacobians. That's a good testbed for adding autodiff support to the code.
• Second metric derivatives. Currently we compute Ricci quantities numerically on the grid because they are so painful to compute analytically for Kerr-Schild. Autodiff might help.
• Linearization of elliptic equations. These are needed for the Newton-Raphson algorithm, but often painful to derive and implement, in particular for complicated equations (XCTS, beyond GR, BNS ID, ...).

[nilsdeppe]

I have played with the library https://github.com/autodiff/autodiff a bit in a toy repo. This seems like it could be useful so we don't have to implement things ourselves. However, I think doing only forward mode autodiff is okay-ish if we have to do it ourselves. I believe we can quite easily extend autodiff to support the blaze SIMD types would should allow autodiff to work.

[nilsvu]

Looks like this library has support for Eigen, so we could model support for Blaze vectors on that.