Limit StaticArray CUDA limited-memory Broyden smoke test

[ChrisRackauckas-Claude]

This PR should be ignored until reviewed by @ChrisRackauckas.

Summary

• Keep scalar CUDA kernel-smoke coverage for SimpleLimitedMemoryBroyden at the default threshold = Val(27).
• Use threshold = Val(2) only for StaticArray SimpleLimitedMemoryBroyden kernel-smoke variants, including the line-search variant.

Context

The latest Sublibrary CI failure was lib/SimpleNonlinearSolve [CUDA], where the self-hosted GPU runner OOMed while launching the SVector{2,Float32} kernel for SimpleLimitedMemoryBroyden{..., Val{27}, ...}. The log reported CUDA memory pool usage of 0 bytes, pointing to CUDA compile/module pressure rather than runtime array allocation.

Validation

• timeout 3600 env NONLINEARSOLVE_TEST_GROUP=CUDA ~/.juliaup/bin/julia +1 --project=. -e 'using Pkg; Pkg.test(; julia_args=["--check-bounds=auto"])' from lib/SimpleNonlinearSolve\n - Passed locally, but both CUDA testsets reported Total 0 because this machine has CUDA.functional() = false and CUDA.has_cuda() = false.\n- ~/.juliaup/bin/julia +1 --startup-file=no -e 'Meta.parseall(read("lib/SimpleNonlinearSolve/test/gpu/cuda_tests__item2.jl", String)); println("parse ok")'\n- git diff --check\n- Runic.format_file(...; inplace=true) on the changed file from a temporary Julia environment with Runic v1.7.0\n

[ChrisRackauckas-Claude]

Scratchpad update: the SimpleNonlinearSolve CUDA sublibrary lane passed in CI (lib/SimpleNonlinearSolve [CUDA], 22m56s), which is the lane targeted by this threshold change. The remaining red checks are unrelated to the CUDA patch: MTK downstream is covered by SciML/.github#110, and the Adjoint resolver failure is covered by the temporary NonlinearSolve guard in #1017 plus the upstream SciMLBase FunctionProperties guard in SciMLBase#1424.