Remove autodiff overhead

[calcmogul]

The autodiff isn't fast enough for embedded environments.

Option 1: Code generation

Generating a linear sequence of stack operations would be faster. It should walk the autodiff tree as usual and write out the C++ expressions instead of computing them. A Doxygen comment should be generated for each function.

Option 2: Expression templates

Since the overhead comes from dynamic dispatch, we could use expression templates instead, though that could get complicated and increase compilation times.

https://en.wikipedia.org/wiki/Expression_templates

[pietroglyph]

I believe I mentioned this and we discussed it on Discord at some point, but it's worth documenting here that something like Enzyme is a good "option 1.5" with similar (or more) advantages to codegen but few of the messy downsides.

[calcmogul]

Yea, you did. Someone in the Faer Discord mentioned Enzyme to me yesterday as well. I'd like to switch to that since autodiff is where most of the solve overhead is; I'm just having a hard time wrapping my head around how to use it in C++, let alone Python.