Allow-list the residual owner-internal names in the sublibrary QA lanes

[ChrisRackauckas-Claude]

This PR should be ignored until reviewed by @ChrisRackauckas.

Follow-up to #4018, which fixed the two systemic QA checks but left 27
sublibrary lanes red. This clears the largest remaining group.

Inventory first

I ran ODEDIFFEQ_TEST_GROUP=QA on Julia 1.12 (the version the QA lanes use)
across all 49 sublibraries locally. Result: 21 pass / 27 fail, matching CI
exactly. The 27 break down as:

Category	Packages	This PR
ExplicitImports on owner-internal names	~18	fixed
public API has docstrings	7	not addressed
Public API documentation errors on 1.12	5	not addressed
No unapproved public reexports	3	not addressed
Aqua (Method ambiguity, Persistent tasks, AllocCheck)	3	not addressed

What this fixes

The Developer Extension API page already states the policy:

Cache types and low-level nonlinear solve helper functions are not public API.

So these names are deliberately non-public and want an allow-list, not a
public declaration. Most of the group is a single omission: OrdinaryDiffEqCore's
precompile workload grew lorenz_pref/lorenz_pref_params, but the ignore lists
that already carry lorenz/lorenz_oop/lorenz_p/lorenz_p_params were never
extended. The remainder:

• Base internals with no public equivalent — structdiff, broadcastable
• names kept in a namespace for dependent sublibraries, whose cross-package use
  ExplicitImports cannot see — _unwrap_val, MatrixOperator, _reshape
  (the same rationale OrdinaryDiffEqCore already documents for _vec,
  _reshape, unwrap_cache)

No check is disabled; the existing per-package ignore lists are extended,
matching the pattern already in these files.

Verification

Every ExplicitImports error across the 27 is gone. 7 lanes now pass fully:

OrdinaryDiffEqBDF            PASS      OrdinaryDiffEqSSPRK          PASS
OrdinaryDiffEqFIRK           PASS      OrdinaryDiffEqTsit5          PASS
OrdinaryDiffEqLowStorageRK   PASS      OrdinaryDiffEqVerner         PASS
OrdinaryDiffEqRKN            PASS

Correction: an earlier revision of this description claimed 9 green lanes and
listed StochasticDiffEqImplicit and StochasticDiffEqLeaping among them. That was
a misread of my own results — both still fail public API has docstrings, which
this PR does not address. Their ExplicitImports errors are fixed (1 → 0 each);
the lanes are not green. CI on this PR caught the discrepancy. The correct count
is 7.

The three that still fail do so only on categories this PR does not target:

OrdinaryDiffEqCore            1 failed  — No unapproved public reexports
OrdinaryDiffEqExponentialRK   1 failed  — public API has docstrings
OrdinaryDiffEqNonlinearSolve  2 failed  — Method ambiguity (Aqua), reexports

Their ExplicitImports errors went 2 → 0, 1 → 0 and 1 → 0 respectively.

Remaining work, for whoever picks it up

• public API has docstrings (ExponentialRK, LowOrderRK, and the SDE family):
  public names with no docstring. Real documentation work.
• Public API documentation errors (5 SDE packages): _has_docstring throws
  Constant binding was imported from multiple modules on Julia 1.12 — a
  binding_module interaction, likely wants an upstream SciMLTesting fix.
• No unapproved public reexports (Core, NonlinearSolve, StochasticDiffEqCore):
  StochasticDiffEqCore exports 22 DiffEqBase-owned names (SDEIntegrator,
  SDEOptions, issplit, …) that DiffEqBase never declared public. That is a
  genuine API-hygiene question — either public declarations in lib/DiffEqBase
  or removal from the export lists — and deliberately not papered over here.
• Aqua: Method ambiguity in NonlinearSolve, Persistent tasks in Default,
  and an AllocCheck failure in AdamsBashforthMoulton (AB3 perform_step!).

🤖 Generated with Claude Code

https://claude.ai/code/session_01TPHRh64BLfoXSzQ3AxKNwC