Raise minimum versions for DowngradeCI

[ChrisRackauckas-Claude]

Please ignore this PR until it has been reviewed by @ChrisRackauckas.

Raise the minimum versions of the symbolic/AD stack used by the test suite so the minimum resolver stack does not produce false @test ... broken unexpected passes in DowngradeCI. The change is compat-only; no public API changed and no deprecation path is required.

The previous minimum stack failed DowngradeCI with unexpected passes at test/lotka_volterra.jl:122 and :123:

• https://github.com/SciML/ModelingToolkitNeuralNets.jl/actions/runs/31891518636/job/95028361788

Regular CI on the current default branch passes:

• https://github.com/SciML/ModelingToolkitNeuralNets.jl/actions/runs/32333763589

Verification

• Exact-minimum pinned stack resolved to ModelingToolkit 11.26.7, ModelingToolkitBase 1.66.0, DiffEqBase 7.15.0, SciMLBase 3.48.0, SciMLSensitivity 7.117.0, DifferentiationInterface 0.7.21, OrdinaryDiffEqVerner 2.3.0, SciMLStructures 1.10.2, Symbolics 7.32.1, SymbolicUtils 4.37.0, and Zygote 0.7.12.
• The exact-minimum Lotka–Volterra gradient diagnostic, with the JET call checks removed only from the diagnostic and the later optimizer section truncated, exited 0; the two Downgrade-sensitive broken comparisons did not produce unexpected passes.
• GROUP=Core julia +1.10 --startup-file=no --project=. -e 'using Pkg; Pkg.test()' ran for one hour and timed out in OptimizationOptimisers.solve with exit 124. No test failure was reported before the timeout; the full optimizer test remains unverified locally.
• typos Project.toml passed.
• git diff --check passed.

This PR targets the main SciML repository; the source branch is hosted on the connected fork only as GitHub’s normal PR mechanism.