Make the QA group scan all four package extensions

[ChrisRackauckas-Claude]

Please ignore this PR until it has been reviewed by @ChrisRackauckas.

Why

SciMLTesting.run_qa runs ExplicitImports' checks on the package module.
ExplicitImports does know about extensions — it reads the [extensions] table out
of Project.toml — but it only checks an extension module that actually exists:

ext_mod = Base.get_extension(mod, Symbol(ext))
ext_mod === nothing && continue

An extension module only comes into existence once its trigger weakdep is loaded, and
the QA environment loaded none of them. Today no PreallocationTools extension is
scanned by QA at all. The fix is to put the weakdeps in test/qa/Project.toml and
using them in test/qa/qa.jl before run_qa.

Coverage

All four weakdeps are pure Julia and resolve cleanly, so all four extensions are now
scanned — nothing is left out:

Extension	Trigger	Now scanned
PreallocationToolsEnzymeCoreExt	EnzymeCore	yes
PreallocationToolsForwardDiffExt	ForwardDiff	yes
PreallocationToolsReverseDiffExt	ReverseDiff	yes
PreallocationToolsSparseConnectivityTracerExt	SparseConnectivityTracer	yes

[compat] entries in test/qa/Project.toml mirror the root package's bounds
(EnzymeCore = "0.8", ForwardDiff = "0.10.38, 1.0.1", ReverseDiff = "1.16",
SparseConnectivityTracer = "1").

Fixed in the extension source (not ignored)

Two findings had a public spelling, so the source was fixed rather than ignored:

• no_implicit_imports — PreallocationToolsForwardDiffExt did bare
  using ForwardDiff / using ArrayInterface / using PrecompileTools and then
  referred to ForwardDiff.Dual, ArrayInterface.restructure, @setup_workload and
  @compile_workload. These are now using ForwardDiff: ForwardDiff,
  using ArrayInterface: ArrayInterface and
  using PrecompileTools: @setup_workload, @compile_workload.
  using Adapt is dropped: the extension never referenced Adapt (it is used only
  in src/PreallocationTools.jl), so keeping it as an explicit import would have
  tripped no_stale_explicit_imports.
• all_qualified_accesses_are_public: Core.apply_type — replaced with
  wrapper{new_parameters...}, which is exactly the same operation spelled with public
  syntax (Core.apply_type is the intrinsic behind T{...}).

Ignore entries added, with justification

Every remaining finding is a name that is genuinely non-public at its owner and has no
public equivalent. Each is grouped under a comment naming the owner in qa.jl.

all_explicit_imports_are_public:

Symbol	Owner	Why unavoidable
:EnzymeRules	EnzymeCore	EnzymeCore neither exports nor declares the submodule public, yet it is the only entry point for defining Enzyme custom rules.
:AbstractTracer, :Dual	SparseConnectivityTracer	SCT exports only TracerSparsityDetector, TracerLocalSparsityDetector, jacobian_sparsity, hessian_sparsity, jacobian_eltype, hessian_eltype, jacobian_buffer, hessian_buffer. These two types are what a get_tmp method must dispatch on for sparsity detection to reach the cache.

all_qualified_accesses_are_public:

Symbol	Owner	Why unavoidable
:forward, :augmented_primal, :reverse	EnzymeCore.EnzymeRules	Declared as bare function ... end stubs and not exported. Adding methods to them is the documented way to write a custom rule.
:Dual, :pickchunksize	ForwardDiff	ForwardDiff exports only DiffResults. Dual is the type every dual-cache method dispatches on; pickchunksize is the chunk heuristic the cache sizing mirrors.
:TrackedArray	ReverseDiff	ReverseDiff also exports only DiffResults; TrackedArray is the type the LazyBufferCache method keys on.
:typename	Base	The only way to recover a DataType's UnionAll wrapper so its type parameters can be substituted. Base offers no public equivalent.

No check was disabled, and no @test_broken / explicit_imports = false escape hatch
was used.

Local verification

GROUP=QA julia --project=. -e 'using Pkg; Pkg.test()' on this branch (Julia 1.12):

Test Summary: | Pass  Total     Time
QA            |   27     27  1m28.9s
     Testing PreallocationTools tests passed

(For reference, the same command with the weakdeps added but before the fixes reported
24 passed, 3 errored, which is the previously-invisible extension debt this PR pays
off.)

The Core.apply_type → wrapper{...} rewrite was additionally checked directly against
replace_type_parameter for Float64, Complex{Float64}, Vector{Complex{Float64}}
and String, all unchanged.

test/qa/Manifest.toml is a local build artifact and is not committed (Manifest.toml
is already in .gitignore). Runic was run on both touched files.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Yb5kCpT5SRzTrhppKSh1n7

[ChrisRackauckas-Claude]

Core group status (informational)

I also ran GROUP=Core locally on this branch, since it touches extension source. It fails, but the identical failure reproduces on unmodified master (ac5aeb9), so it is not caused by this PR:

DiffCache Nested Duals: Error During Test
  LoadError: UndefVarError: `issquare` not defined in `LinearSolveForwardDiffExt`
  Hint: a global variable of this name also exists in SciMLOperators.
ERROR: LoadError: Some tests did not pass: 2 passed, 0 failed, 1 errored, 0 broken.
Test Summary:          | Pass  Error  Total     Time
DiffCache Nested Duals |    2      1      3  2m18.7s

Both runs (master and this branch) produce byte-identical failure output. The root cause is upstream in LinearSolveForwardDiffExt, with LinearSolve v5.3.0 / SciMLOperators v1.25.2 in the resolved manifest. A separate investigation is tracking it; it is out of scope here.

Every other Core testset passed on this branch: Developer Interface (16), DiffCache Dispatch (82 pass / 10 broken), DiffCache ODE tests (7), DiffCache Resizing (64).