Import cache_operator from SciMLOperators in Core tests

[ChrisRackauckas-Claude]

Ignore until reviewed by @ChrisRackauckas.

Summary

• import cache_operator directly from its public owner, SciMLOperators, in the Core test that exercises cached inverse preconditioners
• keep this compatibility fix test-only; there are no production-code changes

Root cause

test/Core/basictests.jl called bare cache_operator while relying on the transitive chain using LinearSolve → reexported SciMLBase names → SciMLOperators.

SciMLBase commit b9c3507a (Make strict QA owner-contract clean (#1472)) removed the blanket SciMLOperators reexport. The dependency boundary was checked directly:

• parent 7b806c50 (SciMLBase 3.40.0): cache_operator is defined after using SciMLBase
• child b9c3507a: it is not
• registered SciMLBase 3.39.1: the full LinearSolve GROUP=Core control passes

SciMLOperators.cache_operator is exported/public, has a docstring, and is included in SciMLOperators' rendered interface documentation.

Verification

• focused cached diagonal-preconditioner regression on SciMLBase 3.40.1: 6/6
• exact Basic Tests on SciMLBase 3.40.1, stacked only for validation with the independent issquare fix in Import issquare from SciMLOperators in ForwardDiff extension #1125: 635/635
• full registered SciMLBase 3.39.1 GROUP=Core: passed
• whole-repository Runic 1.7.0 check: passed
• git diff --check: passed

The full SciMLBase 3.40.1 Core stack continues past this fixed suite and then stops at a separate bare FunctionOperator use in test/Core/default_algs.jl:167. QA on that dependency likewise exposes separate existing owner-contract failures: stale has_concretization in src/LinearSolve.jl:20 and non-owner LinearSolve.SciMLOperators.AbstractSciMLOperator access in ext/LinearSolveSparseArraysExt.jl:58. Those are intentionally outside this focused PR.

[ChrisRackauckas-Claude]

Commit 2589daa7f49a0039d868a6941ed4422ee70c65ba is intentionally limited to the explicit test-owner import in test/Core/basictests.jl.

Local verification plan/result:

1. Reproduce the current SciMLBase 3.40.1 failure on unmodified LinearSolve main.
2. Verify the dependency boundary with registered SciMLBase 3.39.1 and the SciMLBase #1472 parent/child.
3. Verify cache_operator is public and rendered at its owning package.
4. Run the affected 3.40.1 suite with the independent Import issquare from SciMLOperators in ForwardDiff extension #1125 prerequisite, the full registered 3.39.1 Core control, Runic, and git diff --check.

All scoped checks completed as recorded in the PR body. The later FunctionOperator, has_concretization, and qualified-access failures are separate follow-ups and are not being folded into this branch.

[ChrisRackauckas-Claude]

Terminal CI audit: 53 checks passed; the five red checks are independent of this one-line cache_operator owner import, but two earlier classifications need refinement from the completed logs.

• Spellcheck is the clean-main typo isolated in test: spell out SupernodalLU threshold variable #1124.
• SciMLSensitivity is the Distributions resolver boundary isolated in SciMLSensitivity#1575.
• ModelingToolkit reports 11 errors: nine are the DelayDiffEq DEOptions constructor mismatch isolated in OrdinaryDiffEq#4065; the remaining clean-MTK initialization failures are Guess Propagation (missing Initial parameter for x0) and the constant-RHS ForwardDiff y0 is ... not an unknown case, which require their own clean-main investigations.
• QA’s Aqua/ExplicitImports phase actually passed (20 pass, 2 pre-existing broken). The red result is the later JET default-solver test: on the hosted MKL path, BLAS logging calls string(typeof(A)), and JET reports 23 possible runtime dispatches through Base type display. Registered local QA passed, so this environment-specific clean-main failure needs a separate reproduction/bisect; it is not an owner-import failure.
• The Downgrade job has no downloadable log archive. GitHub marks the job failed after about 26 minutes while its julia-runtest step remains recorded as in-progress, so there is not enough evidence to call this a code failure. Separately, local current-dependency Core testing proceeds past this PR and exposes a missing direct test-owner import for FunctionOperator; that is queued as a focused follow-up.

No unrelated changes will be added to #1126.