Fix FieldError for Magnus/Linear integrators with OrdinaryDiffEqDiffe…

[singhharsh1708]

Fix FieldError for Magnus/Linear integrators when OrdinaryDiffEqDifferentiation is loaded

Fixes #3232

Problem

Solving a non-autonomous linear ODE with any Magnus integrator (e.g. MagnusGL6, MagnusGL8, LieEuler, CG2, etc.) crashes with a FieldError when OrdinaryDiffEqDifferentiation is loaded (which happens automatically via DifferentialEquations.jl):

FieldError: type OrdinaryDiffEqLinear.MagnusGL6 has no field `autodiff`,
available fields: `krylov`, `m`, `iop`

Root Cause

MagnusGL6 <: OrdinaryDiffEqLinearExponentialAlgorithm <: OrdinaryDiffEqExponentialAlgorithm{0, false, ...}

In OrdinaryDiffEqDifferentiation/src/alg_utils.jl, the generic dispatch:

function _alg_autodiff(alg::Union{
        OrdinaryDiffEqExponentialAlgorithm{CS, AD},
        OrdinaryDiffEqAdaptiveExponentialAlgorithm{CS, AD},
    }) where {CS, AD}
    return alg.autodiff  # ← crashes: Magnus types have no autodiff field
end

catches all OrdinaryDiffEqLinearExponentialAlgorithm subtypes because they inherit from OrdinaryDiffEqExponentialAlgorithm{0, false, ...}. These algorithms are matrix-exponential methods that never use ForwardDiff/FiniteDiff — they only have fields krylov, m, and iop.

Fix

Add a more-specific _alg_autodiff override that intercepts OrdinaryDiffEqLinearExponentialAlgorithm before the generic ExponentialAlgorithm dispatch:

function _alg_autodiff(::OrdinaryDiffEqLinearExponentialAlgorithm)
    return Val{false}()
end

This is consistent with how OrdinaryDiffEqCore already handles these types in prepare_alg and has_autodiff.

Changes

File	Change
OrdinaryDiffEqDifferentiation.jl	Import OrdinaryDiffEqLinearExponentialAlgorithm from Core
alg_utils.jl	Add _alg_autodiff override returning Val{false}()
differentiation_traits_tests.jl	Regression test verifying _alg_autodiff, prepare_alg, and forwarddiffs_model

Testing

All 3 regression assertions pass with clean precompilation:

✓ _alg_autodiff returns Val{false}()
✓ prepare_alg returns alg unchanged  
✓ forwarddiffs_model returns false

[singhharsh1708]

Replaced the mock struct with real integrators (MagnusGL6, MagnusGL4, MagnusGL8, LieEuler, CG2, CG3, CG4a) and moved the regression test to OrdinaryDiffEqLinear/test/linear_method_tests.jl — no Project.toml changes needed. All 7 pass locally.
Also noticed a pre-existing cG4a convergence order test failure (unrelated to this PR), opened #3466 for it.