Migrate to SciMLTesting v1.2 (off ParallelTestRunner)

[ChrisRackauckas-Claude]

Replaces the ParallelTestRunner-based test/runtests.jl with SciMLTesting v1.2 explicit-args run_tests.

Why

SciML parallelizes tests at the CI-matrix level per GROUP (this repo already uses grouped-tests.yml@v1 and a test/test_groups.toml). ParallelTestRunner's in-process worker parallelism is therefore redundant, and that family of runners is a recurring source of breakage on new Julia releases. This swaps the runner for the shared SciMLTesting harness while keeping the test matrix and exact group semantics.

What changed

• test/runtests.jl: find_tests/runtests + the hand-written GROUP-delete dispatch -> using SciMLTesting; run_tests(; core, qa, umbrellas). The XLA/CUDA memory withenv wrapping is preserved (pinned to one job per process, since each GROUP cell is now its own CI process).
• QA files moved into test/qa/ (qa_tests.jl, doctests.jl) with a new test/qa/Project.toml sub-env. Aqua, Documenter, and ExplicitImports now live in that sub-env and are removed from the root test env.
• Root test/Project.toml: removed ParallelTestRunner; added SciMLTesting and SafeTestsets.
• test/test_groups.toml: unchanged — its Core/QA/GPU keys still drive the CI matrix.

Behavior-preserving group mapping

• Core -> functional suite (models/{fno,deeponet,nomad}_tests, layers/{spectral_conv,spectral_kernel}_tests).
• QA -> Aqua/ExplicitImports + doctests, in the test/qa sub-env.
• GPU -> the self-hosted CUDA cell still runs the full functional + QA suite, expressed as umbrellas = Dict("GPU" => ["Core", "QA"]). The Reactant tests pick up CUDA through reactant_device() as before.
• All (bare local Pkg.test()) -> Core then QA.

Each test file now runs inside its own @safetestset (file-level isolation replaces the old per-file worker isolation). The test bodies and assertions are unchanged.

Verification

• Static: every Project.toml/test_groups.toml parses; all runtests.jl/test .jl files parse with no error nodes; Runic-clean.
• Runtime: the GROUP dispatch (Core / QA / GPU / All) was exercised against SciMLTesting v1.2 locally with recording thunks and routes exactly as the old runtests.jl did (Core->Core, QA->QA-in-subenv, GPU->Core+QA, All->Core+QA). The full ML suite (Reactant/Enzyme/CUDA + a custom Reactant git rev) is left to per-PR CI to validate.

Ignore until reviewed by @ChrisRackauckas.

🤖 Generated with Claude Code