Raise ModelingToolkitBase SymbolicUtils floor to 4.35.3

[ChrisRackauckas-Claude]

Important

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

Summary

• Raise the ModelingToolkitBase SymbolicUtils floor from 4.35.2 to 4.35.3.
• Keep the existing discrete-system @test_nowarn assertion unchanged; it is the focused regression for this dependency-floor failure.
• Make no source-code or test bypass changes.

Cause and boundary

The ModelingToolkit floor change was introduced by SciML/ModelingToolkit.jl commit ab3a1fb, which changed the minimum from 4.34.1 to 4.35.2 while leaving the relevant test unchanged.

SymbolicUtils v4.35.2 still calls deprecated get_canonical_expr(ir, idx) in src/faster_code.jl, including the call reported at line 578 by the downgrade test. JuliaSymbolics/SymbolicUtils.jl commit df29cb30804d4ccacdb35297d3dc07e7c0c4ccea replaces all three calls with get_canonical_expr!. v4.35.3 is the first release tag containing that commit.

No existing ModelingToolkit issue or PR for the 4.35.3 floor was found before opening this PR.

Verification

Clean downgraded reproduction before the change:

• GROUP=InterfaceII with SymbolicUtils 4.35.2 failed the existing Passing nothing to u0 @test_nowarn assertion at lib/ModelingToolkitBase/test/discrete_system.jl:287.
• Captured stderr was the get_canonical_expr deprecation from SymbolicUtils src/faster_code.jl:578.

After raising the floor:

• The exact minimum resolver selected SymbolicUtils 4.35.3, JumpProcesses 9.28.0, and OffsetArrays 1.13.0.
• Full official GROUP=InterfaceII result: 7,235 passed, 0 failed, 2 errored, 4 broken in 50m29.5s.
• Discrete System passed 24/24.
• Implicit Discrete System passed 18/18.
• JumpSystem passed 4,204/4,204.
• After rebasing onto upstream master at 1f408df, the focused original constructor/solve regression passed 2/2, including absence of the deprecated stderr and a successful solve retcode.
• Repository-wide Runic --check . passed after the rebase.
• git diff --check passed.

The two full-group errors are the same independent BoundaryValueDiffEqMIRK ForwardDiff Dual-to-Float64 blocker already tracked in SciML/BoundaryValueDiffEq.jl#556 and addressed by draft SciML/BoundaryValueDiffEq.jl#559. No test was disabled, silenced, or loosened here.

Process

I reproduced the minimum-version workflow locally, inspected the ModelingToolkit compat history and unchanged regression across the introducing commit, inspected the SymbolicUtils commit/tag boundary, reran the official group, rebased onto the latest upstream master, and reran the focused regression plus formatting checks.