ci: revert unsafe sublibrary CI consolidation (restore bespoke per-sublibrary CI; keep faithful main conversions)

[ChrisRackauckas-Claude]

Summary

This PR corrects a CI coverage regression introduced by #951 (CI centralization). An audit of #951 found that consolidating the 10 per-sublibrary CI_*.yml workflows into a single SublibraryCI.yml (a caller of sublibrary-tests.yml@v1) silently dropped test coverage, because sublibrary-tests.yml@v1 defaults to GROUP=Core/GROUP=QA and does not run the real group/tag axes these sublibraries dispatch on.

This PR reverts only the unsafe part of #951 and keeps the faithful parts.

What #951 dropped (the regression)

• NonlinearSolveSciPy: the entire :wrappers suite ran zero items.
• SimpleNonlinearSolve: the :adjoint and :alloc_check groups were no longer run.
• macOS (macos-latest) legs for all 10 sublibraries.
• Per-suite Julia version matrices (e.g. 1.10/1.11/1).
• The main-package CI_NonlinearSolve.yml was converted to tests.yml@v1, which cannot express its local_dependencies (Pkg.develop of the lib/* path deps) or its GROUP test_args.

What this PR does

Reverts (restores bespoke):

• Restores all 10 per-sublibrary CI_*.yml verbatim from before ci: route CI through SciML/.github reusable workflows (@v1) #951, preserving their adjoint/alloc_check/wrappers groups, their ubuntu-latest+macos-latest OS matrices, and their per-suite version matrices.
• Restores CI_NonlinearSolve.yml to its bespoke form (preserving local_dependencies and GROUP test_args).
• Removes the added SublibraryCI.yml.

Keeps (faithful main-package conversions from #951 — these were NOT flagged by the audit):

• Documentation.yml → documentation.yml@v1
• Downstream.yml → downstream.yml@v1 (same 7-package matrix)
• FormatCheck.yml → runic.yml@v1 (Runic preserved)
• SpellCheck.yml → spellcheck.yml@v1
• Downgrade.yml was already a reusable caller (untouched).

Untouched: DowngradeSublibraries.yml (from #950), GPU.yml, TagBot.yml, CompatHelper.yml.

Why the sublibrary CI was intentionally LEFT BESPOKE

sublibrary-tests.yml@v1 auto-discovers lib/* but sends only the default GROUP=Core/QA values. The sublibraries here dispatch their test/runtests.jl on different group/tag names (:wrappers, :adjoint, :alloc_check, plus per-suite OS and version matrices). Consolidating onto the reusable defaults therefore runs zero items for those axes — a silent coverage loss. Per the centralization spec's "MONOREPO SUBLIBRARY CONSOLIDATION — DANGER" rule, the correct action is to leave the per-sublibrary CI_*.yml bespoke rather than convert-and-lose.

Proper follow-up (separate, maintainer-reviewed step)

Adopting sublibrary-tests.yml@v1 safely requires, per sublibrary:

1. Add lib/<name>/test/test_groups.toml encoding the real groups (e.g. core, adjoint, alloc_check, wrappers) plus the OS and Julia-version matrices.
2. Align each lib/<name>/test/runtests.jl GROUP parsing to honor the GROUP value the reusable workflow sends.

Only once both are in place (and verified to run the same item count) should the bespoke CI_*.yml be replaced by SublibraryCI.yml. That work is out of scope here and should be done as its own reviewed PR.

Validation

• All workflow YAML files parse (python3 yaml.safe_load).
• Every remaining reusable caller references a real SciML/.github/.github/workflows/*.yml@v1 workflow, and every input passed is declared by that reusable workflow.
• The net effect on master: the 10 sublibrary CI_*.yml + bespoke CI_NonlinearSolve.yml are restored, SublibraryCI.yml is removed, and the 4 faithful main-package conversions are kept.

---

Please ignore until reviewed by @ChrisRackauckas.

[ChrisRackauckas-Claude]

Superseded by #953: keeping the centralized sublibrary-tests@v1 form and fixing coverage via test_groups.toml + runtests GROUP parsing, instead of reverting to bespoke CI_*.yml.