Run ode_gpu.jl in the GPU test group

[ChrisRackauckas-Claude]

Summary

test/GPU/ode_gpu.jl was added in 2022 (ArrayPartition CuArray / SecondOrderODEProblem discourse fix) but was never included from any test group — git history shows no deliberate disable, only that it was never wired. This adds it to the GPU group body and adds OrdinaryDiffEq (and LinearAlgebra) with compat to test/GPU/Project.toml so the file can resolve.

Verification

This machine has no GPU, so the ODE GPU tests themselves were not executed. Validated only:

1. File parses: julia -e 'Meta.parseall(read("test/GPU/ode_gpu.jl", String))' → OK
2. GPU env resolves in a scratch worktree:

export TMPDIR=/home/crackauc/sandbox/goals/qa-hygiene/jobs/rat-tests/tmp
export JULIA_DEPOT_PATH=/home/crackauc/sandbox/goals/qa-hygiene/jobs/rat-tests/scratch/julia_depot
julia --project=test/GPU -e 'using Pkg; Pkg.instantiate(); Pkg.status()'

Observed status tail:

Status `.../test/GPU/Project.toml`
  Adapt v4.7.1
  ArrayInterface v7.30.2
  CUDA v6.4.0
  KernelAbstractions v0.9.42
  OrdinaryDiffEq v7.8.1
  RecursiveArrayTools v4.5.2 `../..`
  RecursiveArrayToolsArrayPartitionAnyAll v1.0.3 `../../lib/...`
  Zygote v0.7.13
  LinearAlgebra v1.12.0
  Test v1.11.0
GPU_INSTANTIATE_EXIT=0

Runtime correctness of ode_gpu.jl is validated only by the self-hosted GPU CI job (test_groups.toml GPU group on self-hosted / gpu runners).

Not verified

• Actual CUDA execution of ode_gpu.jl (no GPU on this host).
• Interaction with the existing VectorOfArray / ArrayPartition GPU safetestsets beyond env resolve.

Please ignore this draft until reviewed by @ChrisRackauckas.

Risk assessment

• Risk: Low–medium — wires a long-dormant GPU integration smoke test; failure would surface only on self-hosted GPU CI.
• Blast radius: GPU test group only (test/runtests.jl GPU body + test/GPU/Project.toml deps); no library source changes.
• Evidence: Env instantiate + parse succeed locally; no GPU runtime evidence on this host.
• Independent review: Claude Code (Opus 5.5 subagent) rated it low, medium confidence: GPU-group-only change that nobody has run on real CUDA; the env instantiates, the file parses, and the file's steps pass with JLArrays + allowscalar(false). VERDICT MERGE once the self-hosted GPU CI job is green.
• Merge: needs review

Harness: Cursor Agent CLI 2026.09.18-9a7762b; model: auto; transcript: /home/crackauc/sandbox/goals/qa-hygiene/jobs/rat-tests/log.txt on amdci2

Made with Cursor

#662 is fixed by #663 (merged); this PR is its regression test. Rebased onto master to rerun the GPU job.