Speed up compiled state derivatives

[sritchie]

The animation work from the various Clerk related projects has made it clear that the ODE solver can be way faster. The main thing that is is slow is the conversion to and from a nested structure on every call of the derivative function. Instead, we can compile a function that takes a native input and output array of doubles, reads from the input with aget and then mutates the output, returning nothing.

If you want the old behavior, fine, we can pass a flag, or beef up the other compilation step to support it.

For the double pendulum example this gives me at least a 20x speedup.

[sritchie]

So close on this, @littleredcomputer !

[sritchie]

I'm going to call this closed, @littleredcomputer .