refactor: create new LinearProblem instead of remake in ForwardDiff ext

[AayushSabharwal]

No description provided.

[ChrisRackauckas]

What does this actually fix?

[AayushSabharwal]

1. LinearSolve’s init gets LinearProblem with duals in it
2. LinearSolve takes the A and b of that, and makes copies that aren’t duals
3. LinearSolve calls remake with the new A and b
4. remake on a LinearProblem calls SciMLBase.get_new_A_b. This function is responsible for updating the A and b with the appropriate values according to the parameter object, since changing parameters must change A/b
5. As of MTK#master, this errors since p contains duals and it tries to write to A/b with Float64 eltype. I saw this error and thought “oh, I didn’t handle changing the eltype” so if the eltypes are different, I create a similar copy with the correct eltype and return that
6. LinearSolve takes the LinearProblem it got out of remake, expecting the A and b to be Float64 (they’re Duals) and calls init
7. This hits step 1 again

[ChrisRackauckas]

can you make an MWE?