Fix SimpleNonlinearSolve for GPU compatibility

[utkarsh530]

Here are some fixes I made to enable GPU compatibility for SimpleNonlinearSolve.jl. This is required for https://github.com/SciML/ParallelParticleSwarms.jl hybrid algorithms.

• Make SimpleNonlinearSolve algorithms GPU kernel-compatible by eliminating runtime type checks that cause
• TypeError or InvalidIRError during CUDA kernel compilation
• Convert runtime SciMLBase.isinplace() conditionals to @generated functions that resolve at compile time via type parameters
• Standardize NonlinearAliasSpecifier API across all algorithms with a simplified, GPU-compatible pattern
• Centralize shared GPU-compatible helpers (get_alias_u0, should_cache_fx) in Utils module
• Simplify maybe_unaliased to a single method to avoid dynamic dispatch on GPU
• Exclude SimpleHalley from kernel launch tests (requires LU factorization and second derivatives which are fundamentally incompatible with GPU kernels)

Checklist

• Appropriate tests were added
• Any code changes were done in a way that does not break public API
• All documentation related to code changes were updated
• The new code follows the
  contributor guidelines, in particular the SciML Style Guide and
  COLPRAC.
• Any new documentation only uses public API

Additional context

Add any other context about the problem here.

[utkarsh530]

@ChrisRackauckas can you have a look?

[oscardssmith]

Won't this being a generated function have huge compile time impact?

[ChrisRackauckas-Claude]

Review: Unnecessary @generated functions

SciMLBase.isinplace simply extracts a type parameter (iip) and returns it directly:

isinplace(prob::AbstractNonlinearProblem{uType, iip}) where {uType, iip} = iip
isinplace(f::AbstractSciMLFunction{iip}) where {iip} = iip

The compiler constant-folds this to true or false at compile time. @code_typed confirms the function body is literally return false or return true after inference. There is zero runtime work — the dead branch in if SciMLBase.isinplace(prob) gets eliminated by the compiler.

Therefore, none of these needed to be @generated:

1. incompatible_backend_and_problem (NonlinearSolveBase/src/autodiff.jl:108) — reverted to regular function
2. evaluate_f!! (NonlinearSolveBase/src/utils.jl:177) — reverted to regular function
3. evaluate_f (NonlinearSolveBase/src/utils.jl:191) — reverted to regular function
4. should_cache_fx (SimpleNonlinearSolve/src/utils.jl:22) — reverted to @inline function
5. prepare_jacobian (SimpleNonlinearSolve/src/utils.jl:101) — reverted to regular function

Also removed the unnecessary AbstractNonlinearFunction import that was only needed for the @generated pattern.

The @assert in compute_jacobian!! is left commented out — while isinplace infers fine, @assert generates a throw which is genuinely incompatible with GPU kernels (GPUCompiler may not reliably dead-code-eliminate it).

Tests pass:

• NonlinearSolveBase: 16/16 ✓
• SimpleNonlinearSolve: 35,584/35,588 ✓ (4 pre-existing broken)

Fix commit: ChrisRackauckas-Claude/NonlinearSolve.jl@2239d074