remove aliasing from caches

[oscardssmith]

Attempt to drastically reduce the amount of aliasing between different caches to reduce the possibility of getting them out of sync (e.g. in reinit!). Things aren't totally deduplicated yet, but the goal is to only have u, J, fu, and du in a single place each.

[ChrisRackauckas]

The test failures in wrappers looks real

[oscardssmith]

agreed.

[oscardssmith]

@avik-pal

[avik-pal]

Just FYI I will review this. I am trying to get the CI green (in #695) without this patch just so that if know if this actually breaks anything

[avik-pal]

The test failures here are real. After the latest patch we only expect the mac pre tests to intermittently fail

[oscardssmith]

I think I might have just fixed the extension issues. Lets see if CI agrees.

[oscardssmith]

CI seems happy (other than downgrade). @avik-pal this look ready to you? The release strategy here is to release a breaking NonlinearSolveBase change and minor changes for the rest of the libraries. Should this PR include those bumps?

[oscardssmith]

@ChrisRackauckas review addressed. I think this is now good to go.

[ChrisRackauckas]

      [2] (::NonlinearSolveBase.JacobianCache{Float64, SciMLBase.NonlinearFunction{false, SciMLBase.NoSpecialize, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Nothing, Any, Any, Any}, Float64, Float64, ADTypes.AutoForwardDiff{nothing, Nothing}, DifferentiationInterfaceForwardDiffExt.ForwardDiffOneArgDerivativePrep{Tuple{SciMLBase.NonlinearFunction{false, SciMLBase.NoSpecialize, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Nothing, Any, Any, Any}, ADTypes.AutoForwardDiff{nothing, Nothing}, Float64, Tuple{DifferentiationInterface.Constant{Float64}}}, DifferentiationInterfaceForwardDiffExt.ForwardDiffOneArgPushforwardPrep{Tuple{SciMLBase.NonlinearFunction{false, SciMLBase.NoSpecialize, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Nothing, Any, Any, Any}, ADTypes.AutoForwardDiff{nothing, Nothing}, Float64, Tuple{Float64}, Tuple{DifferentiationInterface.Constant{Float64}}}, ForwardDiff.Tag{SciMLBase.NonlinearFunction{false, SciMLBase.NoSpecialize, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Nothing, Any, Any, Any}, Float64}, ForwardDiff.Dual{ForwardDiff.Tag{SciMLBase.NonlinearFunction{false, SciMLBase.NoSpecialize, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Any, Nothing, Any, Any, Any}, Float64}, Float64, 1}, Tuple{Nothing}}}})(::Nothing)

failure looks related.

[oscardssmith]

@ChrisRackauckas I think this is finally ready! The CI error is gone locally.

[ChrisRackauckas]

That test failure is now just lts because lts doesn't use sources.

[oscardssmith]

Agreed. Should I include the version bumps pre merge?

[oscardssmith]

I'll take that as a no, and make a followup for the bumps.

[ChrisRackauckas]

Will these need lower bounds bumps between them?

[oscardssmith]

yes. This is a major bump of NonlinearSolveBase and (I'm pretty sure) a minor bump of everything else (and enforcement of the new NonlinearSolveBase version)