Split sublibrary CI to only test on dependency changes

[ChrisRackauckas-Claude]

Summary

• Splits sublibrary tests out of CI.yml into a new SublibraryCI.yml workflow that only runs tests for sublibraries affected by changed files
• Adds .github/scripts/compute_affected_sublibraries.py which reads the dependency graph from all lib/*/Project.toml files (both [deps] and [extras]), computes the transitive reverse-dependency closure, and maps changed file paths to affected sublibraries
• Global tests always run: Interface, Integrators, AlgConvergence, Regression, Downstream, ModelingToolkit, QA, AD — unchanged in CI.yml

How it works

1. detect-changes job runs git diff and pipes changed files into the Python script
2. Script parses all lib/*/Project.toml to build the internal dependency graph
3. For each changed lib/<pkg>/... file, the script includes that package plus all transitive reverse dependencies
4. The affected list becomes a dynamic GitHub Actions matrix for both functional tests (4 Julia versions) and QA tests (Julia 1 only)

Example behavior

Changed file	Sublibraries tested
lib/OrdinaryDiffEqCore/src/...	All 36 sublibraries
lib/OrdinaryDiffEqTsit5/src/...	Tsit5 + Default + Nordsieck
lib/OrdinaryDiffEqFeagin/src/...	Only Feagin
test/interface/... or src/...	No sublibrary tests (global CI still runs)

Benefits

• Massive CI savings on targeted changes — e.g., a change to OrdinaryDiffEqFeagin only runs 5 jobs (4 versions + QA) instead of ~170
• Self-maintaining — new sublibraries or dependency changes are automatically picked up from Project.toml
• Eliminates ~100 QA exclude entries — QA is now a separate job that only runs on Julia 1

Test plan

• Verify detect-changes job correctly identifies affected sublibraries in CI logs
• Verify sublibrary tests still pass for affected packages
• Verify global tests in CI.yml are unaffected
• Verify that changes only to test/ or src/ (non-lib) don't trigger sublibrary tests

🤖 Generated with Claude Code

[ChrisRackauckas-Claude]

Updated to:

1. Julia script instead of Python — uses stdlib TOML parser, no external deps
2. Sublibraries can declare custom test groups via test/test_groups.toml (e.g., add a GPU group with its own Julia version requirements)
3. Generic group suffix detection in runtests.jl — replaces the hardcoded _QA check with a lookup against actual lib/ directories, so GROUP=OrdinaryDiffEqFoo_GPU correctly resolves
4. Single unified test job — the script outputs the full {group, version} matrix, eliminating the separate test/qa jobs
5. OrdinaryDiffEqCore README change included to trigger all 36 sublibrary tests for verification

[oscardssmith]

This is a very nice change. Good idea.

[ChrisRackauckas]

This was needed to complete the StochasticDiffEq and DelayDiffEq stuff haha.