compat: restore CUDA v5 lower bound

[ChrisRackauckas-Claude]

Note

Draft PR — please ignore until reviewed by @ChrisRackauckas.

Closes #449.

Why

PR #441 cleanup bumped CUDA = "5, 6" → "6" along with a batch of other major-version drops. The CUDA part was unnecessarily aggressive: CUDA 6 was released with no breaking changes from 5 (see v6.0.0 release notes), and the lift broke downstream consumers like NeuralLyapunov whose stack still requires CUDA 5 transitively (via AdvancedHMC, see SciML/NeuralLyapunov.jl#175).

What

CUDA = "6" → CUDA = "5, 6" in:

• Project.toml
• test/Project.toml
• docs/Project.toml

No code changes. The CUDAExt surface is tiny:

# ext/CUDAExt.jl
DiffEqGPU.maxthreads(::CUDABackend) = 256
DiffEqGPU.maybe_prefer_blocks(::CUDABackend) = CUDABackend(; prefer_blocks = true)
function DiffEqGPU.lufact!(::CUDABackend, W)
    CUDA.CUBLAS.getrf_strided_batched!(W, false)
    return nothing
end

CUDA.CUDABackend and CUDA.CUBLAS.getrf_strided_batched! are both stable across CUDA 5 and 6, so no shim is needed. Test code uses only CUDA.CUDABackend() (in test/utils.jl).

Verification

Local resolution checks (Julia 1.10, no GPU on the test machine — CI will exercise the actual GPU paths):

# CUDA = "5, 6" in package + test envs
julia --project=test -e 'using Pkg; Pkg.add(PackageSpec(name="CUDA", version="5"))'
# → resolves to CUDA v5.11.2; DiffEqGPU.CUDAExt + LinearSolveCUDAExt + DiffEqBaseCUDAExt all precompile

julia --project=test -e 'using Pkg; Pkg.add(PackageSpec(name="CUDA", version="6"))'
# → resolves to CUDA v6.1.0; same extensions all precompile

Both endpoints of the new range resolve cleanly with the rest of the test stack (OrdinaryDiffEq 7, ModelingToolkit 11, SciMLBase 3, etc.).