Support NonlinearSolve.jl for the ODE nonlinear solver

[oscardssmith]

This is WIP, but it mostly works (for in place ODE only), (and is a lot slower since it's missing all the fancy optimizations still).

[oscardssmith]

so the current status is that out of place solves now work (although something is going wrong since it is getting a lot of nonlinear solver convergence failures). In place is broken because _compute_rhs! isn't autodiff compatible, and no matter what I do, NonlinearSolve is trying to autodiff through it.

[avik-pal]

In place is broken because _compute_rhs! isn't autodiff compatible

How does the default newton in ordinarydiffeq handle this?

and is a lot slower since it's missing all the fancy optimizations still

you probably need the recompute_jacobian option here to mimic the jacobian reuse behavior https://github.com/SciML/NonlinearSolve.jl/blob/0dfc1358be16478a29e8940b5abca896d30ec9ff/src/core/generalized_first_order.jl#L204-L205

[oscardssmith]

with recompute_jacobian=false this actually is looking really good! I'm getting better stats than NLNewton (admitadly so far I've only been testing on f(u,p,t)=u. That said, I do seem to have somehow messed upNLNewton` in the process...

[oscardssmith]

it is worth mentioning that it is very possible that the better stats come from me not updating the stats properly. Still needs testing.

[ChrisRackauckas]

It's missing tests?

[oscardssmith]

I've added some tests that use this. I'm not sure how much needs testing.

[ChrisRackauckas]

Match a bit of what NLAnderson has done https://github.com/search?q=repo%3ASciML%2FOrdinaryDiffEq.jl+NLAnderson&type=code

[avik-pal]

Are we using the jacobian reuse here? or is that a plan for later

[oscardssmith]

I believe everything should now work. The last issues were that tstep and invγdt were being computed correctly, but the incorrect values for them were being passed into the nlp_params, leading to all sorts of nonsense.