Configure the sublibrary QA lanes for monorepo docs and facade reexports

[ChrisRackauckas-Claude]

This PR should be ignored until reviewed by @ChrisRackauckas.

The failure

Every lib/* QA lane is red — ~45 sublibrary-ci / lib/… [QA] jobs. This is
easy to miss because the sublibrary matrix is path-filtered: a PR that does not
touch lib/ spawns zero sublibrary QA jobs, so its Sublibrary CI goes green
vacuously. #4014 touched lib/OrdinaryDiffEqCore/Project.toml, which triggered
the full matrix and surfaced it.

To be clear about causation: this pre-dates #4014. I checked out
eb3888cf7 (before that merge) and ran the Feagin QA lane directly:

$ ODEDIFFEQ_TEST_GROUP=QA julia +1.12 --project=lib/OrdinaryDiffEqFeagin -e 'using Pkg; Pkg.test()'
public API is rendered in docs: Test Failed
No unapproved public reexports: Test Failed
ERROR: LoadError: Some tests did not pass: 19 passed, 2 failed, 0 errored, 0 broken.

Why

SciMLTesting 2.x defaults api_docs and check_reexports to true. Both
assumptions are wrong for solver sublibraries:

1. Rendered docs. run_api_docs resolves docs_src to <pkgroot>/docs/src,
which no sublibrary has — lib/OrdinaryDiffEqFeagin/ contains only
src/ test/ Project.toml README.md LICENSE.md. A missing directory yields an
empty rendered set, so every public name is reported unrendered:

┌ Info: run_api_docs: public API names not rendered in a @docs block
│   pkg = OrdinaryDiffEqFeagin
│   docs_src = ".../lib/OrdinaryDiffEqFeagin/docs/src"
│   unrendered = 182-element Vector{Symbol}

These packages are documented in the umbrella manual, where the root QA lane
already checks rendering.

2. Facade reexports. The solver packages @reexport using SciMLBase (or
StochasticDiffEqCore / DiffEqBase) by design — several qa.jl files already
say so in a comment. With no reexports_allow, that deliberate surface is
reported wholesale (185 names for Feagin).

The fix

Per package, mirroring what #3988 did for the umbrella:

api_docs_kwargs = (; rendered = false),
reexports_allow = union(public_api_names(SciMLBase), (:SciMLBase,)),

Neither check is switched off:

• the docstring half of the api-docs check stays on, so a public name with
  no docstring is still caught;
• reexports from any package other than the deliberately-reexported one are
  still reported.

Mapping is mechanical, from each package's own @reexport using:
SciMLBase (30), StochasticDiffEqCore (9), DiffEqBase (2), no reexport (7 —
rendered = false only). The module name is already in scope via the package's
own reexport, so no test-environment or Project.toml changes are needed.
GlobalDiffEq and ImplicitDiscreteSolve already point docs_src at the monorepo
manual; their api-docs configuration is left alone.

Verification

Ran ODEDIFFEQ_TEST_GROUP=QA on Julia 1.12 (the version the QA lanes use),
covering every shape in the change:

Package	Group	Result
OrdinaryDiffEqFeagin	SciMLBase	tests passed
OrdinaryDiffEqExplicitRK	SciMLBase, pre-existing api_docs_kwargs merged into	tests passed
OrdinaryDiffEqDifferentiation	no reexport	tests passed
GlobalDiffEq	DiffEqBase, bespoke docs_src preserved	tests passed
ImplicitDiscreteSolve	SciMLBase, bespoke docs_src preserved	tests passed

What this does not fix

Some sublibraries stay red on a separate, pre-existing seam unrelated to
these two checks. In each case the two checks this PR targets do now pass:

• OrdinaryDiffEqTsit5 — Public API documentation 1 pass, No unapproved public reexports 1 pass; still errors on ExplicitImports
  (all_qualified_accesses_are_public).
• StochasticDiffEqMilstein — reexports check passes; Public API documentation still errors inside the docstring half with
  Constant binding was imported from multiple modules (a Julia 1.12
  binding_module interaction, not the rendered check).
• OrdinaryDiffEqCore — rendered check passes; still has a stale explicit
  import (_unwrap_val) and a non-public qualified access.
• StochasticDiffEqCore — still reports 22 reexported names (SDEIntegrator,
  SDEOptions, issplit, …). These are DiffEqBase-owned names the sublibrary
  exports even though DiffEqBase never declared them public. That is a real
  API-hygiene finding, so I deliberately did not allowlist it away — it
  wants either public declarations in lib/DiffEqBase or removal from the
  sublibrary export lists, as its own change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TPHRh64BLfoXSzQ3AxKNwC