Recursive parallelism in evaluation

[MilesCranmer]

This splits the trees into threads when evaluating them. It also has a mechanism to return early if only one of the two threads hits a NaN. Thus we should detect and return early for NaNs much quicker.

It does this by using a shared channel that the threads can post to. It has a spin lock on this shared channel.

TODO:

• [ ] Find a way to kill a thread early if returning
• [ ] Fix issue with unstable type return value.
• [x] Check if Enzyme still works with this on.
  • Nope!
• [ ] Disable this by default (as it is experimental), and have a flag for turning it on.
• [ ] See if this is faster at all compared to regular evaluation.
• [x] Fix scoping issue.

[MilesCranmer]

Probably not worth it right now.