qa: import names through owning modules in ModelingToolkitBase

[ChrisRackauckas-Claude]

What and why

lib/ModelingToolkitBase still used the blanket-using import style that src/ModelingToolkit.jl was already migrated away from, so all six of its ExplicitImports checks errored, and its re-export check had no approved list. This is the debt explicitly deferred in #4934 ("The full strict QA still has separate baseline JET, ExplicitImports, and unapproved-reexport debt. This PR does not suppress or allowlist those failures.").

What was wrong, concretely:

• 200 implicitly-imported names (bare using SciMLBase, using LinearAlgebra, using Graphs, …)
• 42 stale explicit imports (Chain/Postwalk/Prewalk/Fixpoint, BuildTargets/JuliaTarget, gradient/hessian/derivative/sparsehessian, Base: RefValue, ADTypes: AutoForwardDiff, …)
• 8 explicit imports and 13 qualified accesses reaching a name through a re-exporter rather than its owner
• the _EvalModuleOpt* submodules implicitly importing RuntimeGeneratedFunctions
• Moshi @data-generated submodules (MissingGuessValue, StructuralHint) being unanalyzable
• public_reexports reporting all 145 externally-owned public names, because run_qa was called without reexports_allow

This imports each name from the module that defines it, deletes the stale ones, and adds the curated lists that test/qa/qa.jl already carries for ModelingToolkit — NONPUBLIC_EXPLICIT_IMPORTS (27), NONPUBLIC_QUALIFIED_ACCESSES (96), REEXPORTED_API (145), and allow_unanalyzable for the two Moshi modules. No check is disabled and no wildcard is used; every entry names something genuinely owned elsewhere.

Dependency changes

TermInterface, StaticArraysCore and IntervalSets become direct deps of ModelingToolkitBase, since they own names it already uses (maketerm/metadata, SVector/MVector, endpoints) that were previously reached through re-exporters. All three were already transitive dependencies, and all three are MIT — same as ModelingToolkitBase — so this adds no new licensing surface.

ADTypes moves from [deps] to test-only: its source use was dead, and leaving it in [deps] would fail Aqua's stale-dependency check.

Verification

check	result
GROUP=QA (ModelingToolkitBase)	ExplicitImports 6/6 pass, re-exports pass
GROUP=InterfaceI (ModelingToolkitBase)	InterfaceI | 1599 pass, 5 broken, 1604 total — passed
GROUP=QA (umbrella ModelingToolkit)	QA | 52 pass, 52 total — passed; exercises @import_mtkbase, which pulls in every MTKBase name
load + build a problem	using ModelingToolkit, @mtkcompile, ODEProblem → u0 = [1.0]
runic --check over lib/, src/, test/	exit 0
typos	exit 0

The QA group goes from 13 pass / 2 fail / 6 error to 20 pass / 1 fail / 0 error.

This does NOT make the QA job green — the JET lane still fails

The one remaining failure is JET-test failed, ═════ 263 possible errors found ═════. It is unchanged by this PR: the findings were extracted before and after, gensym counters normalised, and the diff is empty (263 lines either way).

Of the 263, 249 are local variable ##And#NNN#f#1 may be undefined / ##Call#NNN#... — JET's typo mode misreading the short-circuit guards that Moshi.Match.@match lowers to, the same class already tracked in #4958, JuliaLang/julia#62745 and aviatesk/JET.jl#858. The remaining 14 are is not defined findings that look genuine and are separate bugs (convert_system ×2, get_parameter_dependencies, ModelingtoolkitizeParametersNotSupportedError, field accesses .v/.sz/.s/.new/.expr, locals newbuf ×3, dsti, buffer).

test/qa/qa.jl already sets jet = false for ModelingToolkit with a comment citing those three issues. Applying the same treatment to lib/ModelingToolkitBase/test/qa/aqua.jl would be a one-line change — but that is disabling a failing check, so I have deliberately not done it. Whether to mirror that precedent here is a maintainer call, and the 14 real findings probably deserve their own issue either way.

Scope

Deliberately kept to import hygiene and the QA lists — no behavior changes. Split out of #4983, where these failures surfaced, so that a mechanical sweep does not ride along with a behavior change.

Not verified

GROUP=InterfaceII, the docs build, downstream and downgrade lanes, and non-Linux or non-1.12 Julia versions were not run locally.

---

Please ignore until reviewed by @ChrisRackauckas.