_init_originator_gradient: wrap nested Enzyme.gradient with set_runtime_activity

[ChrisRackauckas-Claude]

Please ignore this PR until reviewed by @ChrisRackauckas.

Summary

The EnzymeOriginator overload of _init_originator_gradient (src/concrete_solve.jl:377) called

Enzyme.gradient(Enzyme.Reverse, Enzyme.Const(f), tunables)

without set_runtime_activity. The outer user-level Enzyme.autodiff(set_runtime_activity(Reverse), ...) call's runtime-activity relaxation does not propagate into this nested gradient, so the inner Enzyme call raised EnzymeRuntimeActivityError at MTK's init remake even when the caller used the documented pattern (see PR #1455).

This patch wraps the inner gradient with set_runtime_activity(Reverse) so the nested gradient inherits the same activity relaxation as the outer call.

Diff

function _init_originator_gradient(::SciMLBase.EnzymeOriginator, f, tunables)
    return Enzyme.gradient(
        Enzyme.set_runtime_activity(Enzyme.Reverse),
        Enzyme.Const(f), tunables,
    )[1]
end

Effect on #1455

After this patch the EnzymeRuntimeActivityError is gone, but the @test_broken Enzyme block in test/parameter_initialization.jl still trips a different downstream error: MethodError: no method matching MixedDuplicated(::ODESolution{...}) inside Enzyme's runtime-activity wrapping of the ODESolution type. That is a separate Enzyme/MTK problem unrelated to the inner-gradient-activity issue this PR fixes. So #1455 stays @test_broken for now; this PR is independently useful for anyone hitting the runtime-activity error from a clean Enzyme call path.

Test plan

• CI green on master.
• Reproduction of test/parameter_initialization: use Enzyme.autodiff with Duplicated #1455's Enzyme block locally now fails with the MixedDuplicated/ODESolution MethodError rather than EnzymeRuntimeActivityError, confirming this PR moves past the inner-activity wall.