Wire allocation tests into the CPU group

[ChrisRackauckas-Claude]

Summary

test/alloc_tests.jl was added in e72e1aa gated behind GROUP == "all" || GROUP == "nopre", but neither group exists in test/test_groups.toml or in any workflow — so the tests have never run in CI. This moves the safetestset into the GROUP == "CPU" block (which also covers the "CPU 32-bit" CI group, group = "CPU") and removes the dead gate.

Second commit fixes a Julia-version sensitivity in the test itself: on Julia <= 1.10, @allocated evaluated directly in a @testset body also counts boxing the measured call's result, so the three == 0 checks fail on lts (32 == 0 for lorenz, 16 == 0 for diffeqgpunorm) even though the code under test does not allocate. The measurements are now made inside zero-arg functions, which report 0 on both 1.10 and 1.13 and still detect real allocations. No VERSION gate and no loosened thresholds.

Verification

Machine has no NVIDIA GPU; verified on CPU only. Envs were resolved per Julia version from test/Project.toml with DiffEqGPU Pkg.develop'd at this checkout (not the registered release).

Safetestset wired exactly as runtests.jl includes it, on Julia 1.10.12 before the fix — reproduces the CI failure:

Test Summary:                                              | Pass  Fail  Total   Time
Allocation Tests                                           |    1     3      4  25.2s
  User-defined function should not allocate with SVector   |          1      1   1.8s   (32 == 0)
  make_prob_compatible should be low allocation            |    1            1   4.8s
  diffeqgpunorm should not allocate for SVector            |          1      1   0.2s   (16 == 0)
  diffeqgpunorm should not allocate for scalars            |          1      1   0.0s   (16 == 0)
ERROR: LoadError: Some tests did not pass: 1 passed, 3 failed, 0 errored, 0 broken.

Root cause probe on 1.10.12 (same calls, different measurement site):

testset-scope lorenz:        32      testset-scope gpunorm:     16
function-wrapped lorenz:      0      function-wrapped gpunorm:   0
escaping-allocation control through a wrapper (writes [x,x] into a Ref): 64  -- still detected

After the fix, GROUP=CPU julia +1.10 --project=<dev-env> test/runtests.jl (1.10.12) — full CPU group green, including:

Test Summary:                 | Pass  Total   Time
Adaptive endpoint termination |  132    132  37.1s
Test Summary:    | Pass  Total  Time
Allocation Tests |    4      4  0.2s

GROUP=CPU julia +1 --project=<dev-env> test/runtests.jl (1.13.1) — full CPU group green, tail:

Test Summary:                 | Pass  Total   Time
Adaptive endpoint termination |  132    132  43.7s
Test Summary:    | Pass  Total  Time
Allocation Tests |    4      4  0.2s

No Fail/Error/did not pass entries in either log. runic --inplace made no changes and typos is clean on both touched files.

Not verified / notes

• test/reverse_ad_tests.jl stays commented out. It was disabled in c550bb9 ("test without reversediff", Fix master tests (v1.10 and v1.11) #348) during the SciMLBase v3-era test fixes, and it still fails today: Zygote's pullback hits SciMLBaseChainRulesCoreExt's rrule for ODESolution, which constructs the solution with an outdated argument list — an upstream SciMLBase/ChainRules issue, not something this repo's test wiring can fix.
• The tests / CUDA job failure seen in CI (Enzyme ensemble gradients, test/enzyme/cuda_records.jl) is pre-existing on master — tracked in CUDA CI: test/enzyme/cuda_records.jl fails on released CUDACore (Enzyme undef-shadow fill! for non-numeric eltypes) #546 — and unrelated to this diff. No GPU paths were run locally.
• 32-bit: not run locally; the measurement is now inside functions so the previously observed 1.10-vs-1.13 difference does not apply, and the thresholds are == 0 byte counts / < 1024 on host-side code.
• Julia 1.11 is not in the CI matrix and was not run.

Please ignore this draft until reviewed by @ChrisRackauckas.

Risk assessment

• Risk: Low — adds 4 host-side @allocated assertions (~1s) to the CPU group; no src changes.
• Blast radius: test/runtests.jl dispatch plus the measurement pattern in test/alloc_tests.jl; the CPU 32-bit job also runs it.
• Evidence: full GROUP=CPU run of test/runtests.jl observed green locally on Julia 1.10.12 and 1.13.1 against the dev'd checkout, including the newly wired safetestset; red-before/green-after for the measurement fix.
• Independent review: pending
• Merge: needs review

🤖 Generated with Devin CLI 3000.11.3 (model: swe-2-high). Local session, transcript at /home/crackauc/sandbox/goals/qa-hygiene/jobs/gpu-fixes/log.txt on amdci2.