Fix mass-matrix Rosenbrock test (Float32 dt) and raise downgrade compat floors

[ChrisRackauckas-Claude]

Problem

On master, the CPU / JLArrays / OpenCL / QA test groups (all Julia versions) error on the first test, test/gpu_kernel_de/stiff_ode/gpu_ode_mass_matrix.jl, so every group goes red before reaching anything else:

MethodError: Cannot `convert` an object of type
  StaticWOperator{true, SMatrix{3,3,Float64,9}, ...} to
  StaticWOperator{true, SMatrix{3,3,Float32,9}, ...}
  @ setproperty!(::JacReuseState{Float64, SMatrix{3,3,Float32,9}, ...}, ::Symbol, ::StaticWOperator{...Float64...})
  @ OrdinaryDiffEqDifferentiation derivative_utils.jl:922  (jac_reuse.cached_W = W)

Root cause

The reference CPU bench solve used dt = 0.1 (Float64) on an otherwise all-Float32 problem (u0, p, tspan, tolerances are Float32; the GPU solve already uses dt = 0.1f0). OrdinaryDiffEqRosenbrock's JacReuseState (added in Rosenbrock 1.31.1 / all 2.x) types its cached_W slot from the state eltype (Float32), but with a Float64 dt, calc_W builds a Float64 StaticWOperator that cannot be stored into the Float32 slot. This is a genuine upstream mixed-precision bug, surfaced here by the test mixing a Float64 dt into a Float32 problem.

Fix

1. Mass-matrix test: use a type-consistent dt = 0.1f0 for the reference bench solve (matches the Float32 problem and the GPU solve). The GPU-vs-reference tolerance assertions are unchanged — nothing is loosened, skipped, or broken.

2. Downgrade compat floors (the Downgrade job hit a ModelingToolkit unsatisfiability: at downgrade StaticArrays was pinned to its 1.9 floor, forcing MTK below v11 while the test pins MTK ≥ 11.17):

   • StaticArrays 1.9 → 1.9.14 (MTK 10.18–11 requires StaticArrays ≥ 1.9.14)
   • SciMLBase 3 → 3.1 (lowest MTK compatible with SciMLBase 3 is 11.22; JumpProcesses 9.25.1+ requires SciMLBase ≥ 3.1.0, excluding 3.0.0)
   • ModelingToolkit 11.17.0 → 11.23.0 (test; lowest MTK consistent with the SciMLBase 3.1 and OrdinaryDiffEqRosenbrock 2 floors)

Verification (local)

• Julia 1.12, GROUP=CPU: GPU Kernelized Stiff ODE Mass Matrix safetestset passes 2/2.
• Julia 1.10: the floor set resolves cleanly (SciMLBase 3.1.0 / MTK 11.23.0 / StaticArrays 1.9.14 / OrdinaryDiffEqRosenbrock 2.0.0).
• Normal latest-version resolution is unaffected (resolves SciMLBase 3.21, MTK 11.28, StaticArrays 1.9.18).
• Changed .jl is Runic-clean.

Notes / out of scope

• The Documentation and CUDA jobs run on self-hosted GPU runners and could not be reproduced locally. Documentation additionally fails on GPU-only issues (an AD DimensionMismatch, Rodas5 no longer auto-exported by OrdinaryDiffEq v7, and a new ModelingToolkit problem representation that is not GPU-inline-allocatable) that are separate from this PR.
• The Downgrade job may still hit the pre-existing, maintainer-noted ForwardDiff/LogExpFunctions strict-floor wall (an upstream issue tracked separately); this PR removes the current MTK blocker so the downgrade can get further.

Please ignore until reviewed by @ChrisRackauckas