Align NLStep TRBDF2 parity test estimator

[ChrisRackauckas-Claude]

Ignore this PR until it has been reviewed by @ChrisRackauckas.

Summary

• compare the standard and nlstep = true TRBDF2 paths with smooth_est = false
• keep the existing 2e-4 saved-trajectory threshold unchanged
• leave solver behavior unchanged; this only restores the intended parity-test semantics

Cause

The smoothed-estimator change in #3823 lets the full-state NonlinearSolveAlg path reuse its W solve, but the reduced MTK NLStep problem intentionally cannot provide that full-state solve and falls back to the raw estimator. The parity test therefore started comparing two different error-estimation strategies rather than only the two nonlinear-solve paths.

An exact local diagnostic gave a maximum saved-trajectory delta of 3.052191084731426e-4 with the default mixed estimators and 1.2863503533355747e-4 when both paths used the raw estimator. The latter satisfies the existing threshold without weakening it.

Smoothed-error behavior remains covered separately by nsa_smooth_est_tests.jl.

Local verification

• GROUP=ModelingToolkit julia +1.12 --project=lib/OrdinaryDiffEqNonlinearSolve -e 'using Pkg; Pkg.test()' on the exact failing parent: NLStep 23 pass / 2 pre-existing broken; preconditioners 18/18; DAE initialization 21/21
• the same official group after rebasing onto current master (b22e2e58b2), with clean local SciMLBase 3.40.1 required by current compat: the same fully passing results
• Runic 1.7.0 passes on the changed file; git diff --check passes

Whole-repository Runic on clean current master has three unrelated pre-existing formatting failures. They are reproduced independently and already fixed by draft PR #4064.

[ChrisRackauckas-Claude]

Hosted CI classification after the first run:

• The relevant OrdinaryDiffEqNonlinearSolve [ModelingToolkit] job never reaches the test command. julia-buildpkg fails because OrdinaryDiffEqCore requires SciMLBase 3.40, while General currently provides versions only through 3.39.1.
• The root ModelingToolkit test fails during resolution for the same reason: OrdinaryDiffEqNonlinearSolve requires SciMLBase 3.40, while OrdinaryDiffEq resolves through 3.39.1.
• The downstream ModelingToolkit Initialization job also fails during resolution because the developed DiffEqBase requires SciMLBase 3.40 and the registry stops at 3.39.1.
• Runic explicitly passes the changed lib/OrdinaryDiffEqNonlinearSolve/test/modelingtoolkit/nlstep_tests.jl. It fails only on the three pre-existing DISCO formatting files already handled by Run Runic on the disco-optimizations files #4064.

These are clean-base/external-registration blockers, not failures introduced by this test-only commit, so I have not changed the branch in response. The exact relevant group passes locally on current master when developed against clean local SciMLBase 3.40.1, as recorded in the PR body.