Avoid piracy in jump dependency traversal

[ChrisRackauckas-Claude]

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

Summary

• keep jump-specific dependency traversal behind a ModelingToolkitBase-owned helper
• stop defining SymbolicUtils.search_variables! methods on JumpProcesses-owned jump types
• add a regression test that rejects those cross-package methods

Context

A clean origin/master QA reproduction at 731eff66b596fbc6bc36aff45b21bb2e81e1bddb reports 14 Aqua piracy methods: four search_variables! methods and ten older toexpr methods. This focused change removes the four jump-related methods. The remaining toexpr ownership work is tracked in #4670.

The existing narrow SimpleNonlinearSolve stale-dependency allowance is unchanged. That dependency is used only by MTKBifurcationKitExt and must remain hard so loading BifurcationKit activates the extension.

Local verification

• Runic --check --diff on all three changed files: pass
• git diff --check: pass
• focused lib/ModelingToolkitBase/test/dep_graphs.jl: all testsets pass (1/1, 22/22, 3/3, 1/1)
• full GROUP=InterfaceI package test: 1,502 pass, 5 pre-existing broken, 1,507 total; package test passed in 48m04s
• Aqua.test_piracies(ModelingToolkitBase): expected nonzero result containing exactly the ten pre-existing toexpr methods and no search_variables! methods

CI baseline

The repository-wide Runic check fails on unchanged clean-master lib/ModelingToolkitBase/src/problems/initializationproblem.jl. Exact local reproduction and the introducing commit are tracked in #4908. This PR does not mix that unrelated one-word formatting fix into the jump-piracy change.

[ChrisRackauckas-Claude]

CI baseline update: the two completed failures are both reproduced on unchanged master and are unrelated to this patch.

• repository-wide Runic: one missing formatter-required return in initializationproblem.jl, tracked with the introducing commit in Clean master Runic fails on InitializationProblem return #4908
• ModelingToolkitStandardLibrary.jl/Core/1/: identical current-master resolver conflict between DataInterpolations 8 and SciMLBase 3.35+ / FindFirstFunctions 3, tracked in Test resolver conflict: DataInterpolations 8 vs SciMLBase 3.35+ ModelingToolkitStandardLibrary.jl#498

The PR-specific jump dependency tests and full local InterfaceI group pass; no unrelated compatibility or formatting changes are included here.

[ChrisRackauckas-Claude]

Additional clean-master CI baseline: the LTS benchmark job fails because benchmark/benchmarks.jl relies on the removed OrdinaryDiffEqDefault blanket SciMLBase reexport and does not import SciMLBase.init. Clean Julia 1.10 namespace reproduction and the introducing OrdinaryDiffEqDefault change are tracked in #4910. The Julia 1 benchmark cell was canceled by the matrix after the LTS cell failed.

[ChrisRackauckas-Claude]

Additional clean-master CI classification: the MTKBase compatibility-floor failure is identical on origin/master and is tracked separately in #4911. History analysis traces the stale Julia 1.10 @test_broken @inferred expectation to d893f4205a56dc1bdfdb729661cdbd03c4ae6da7; it became the active downgrade failure after 84dc0e3f2644b3a2e49bb213d869c129ceb41be9 lowered the Moshi floor and made both inference checks pass. No downgrade-related changes are included in this focused PR.