Use SciMLOperators as the operator API owner

[ChrisRackauckas-Claude]

Ignore until reviewed by @ChrisRackauckas.

Summary

• remove the stale root has_concretization import
• import AbstractSciMLOperator only through its owner, SciMLOperators
• use the already imported owner type in LinearSolveSparseArraysExt
• remove unused non-owner AbstractSciMLOperator imports from the CUDA and Metal extensions

Dependency boundary

SciMLBase commit b9c3507a (#1472) removed @reexport using SciMLOperators. A fresh current-main checkout with local SciMLBase 3.40.1 reproduced exactly two ExplicitImports errors: the stale has_concretization root import and LinearSolve.SciMLOperators.AbstractSciMLOperator accessed through a non-owner in the sparse extension.

SciMLOperators 1.25.1 is the declaring owner. Runtime checks confirm AbstractSciMLOperator is public and has_concretization is exported; both have docstrings and rendered docs/src/interface.md entries.

Verification

• unchanged main, local SciMLBase 3.40.1: reproduced 18 passes, 2 ExplicitImports errors, and 2 existing broken cases
• fixed GROUP=QA, local SciMLBase 3.40.1: passed (110 passes, 10 existing broken cases across QA/JET; allocation sets 48/48 and 8/8)
• fixed GROUP=QA, registered-current SciMLBase 3.39.1: passed with the same counts
• owner-clean GROUP=Core: 1,580 top-level assertions passed before the independent current-main FunctionOperator error owned by draft Import FunctionOperator from SciMLOperators in Core tests #1127; relevant sparse sets passed, including re-solve 133/133, SupernodalLU 72/72, sparse-vector RHS 12/12, and nonstructural zeros 141/141
• official GPU environment: LinearSolveCUDAExt compiled and loaded successfully; no GPU runtime test was claimed on this host
• Runic on all changed files and the full repository: passed
• git diff --check: passed

Independent current-main prerequisites

• Import FunctionOperator from SciMLOperators in Core tests #1127 owns the FunctionOperator Core failure exposed by SciMLBase 3.40.1
• test: spell out SupernodalLU threshold variable #1124 owns the two unchanged thr spell-check findings on current main

No MKL/JET changes or QA suppressions are included.

[ChrisRackauckas-Claude]

One-hour terminal CI audit for exact head ac588f7beb: 55 checks succeeded, 4 were skipped, and 3 failed. All three are independent current-main failures; the owner-import patch itself has no hosted test failure.

• Typos: exact unchanged findings at test/Core/supernodal_lu.jl:176-177 for thr, owned by focused draft test: spell out SupernodalLU threshold variable #1124.
• SciMLSensitivity/Core1: setup resolves registered SciMLBase 3.39.1, then source SciMLSensitivity requires 3.40.0-3; resolution stops with no versions left, before tests. The prerequisite General compatibility transition in Retroactively cap SciMLBase at 3.39 for packages that lose reexported names JuliaRegistries/General#162635 is still open.
• ModelingToolkit/All: exact result is 758 passed, 11 errors, 14 pre-existing broken cases. The same job against exact clean LinearSolve main reproduces the identical 758/11/14 result with the same dependency graph. Nine errors are the DelayDiffEq compatibility/runtime set addressed by separately tested Retroactively cap DelayDiffEq at OrdinaryDiffEqCore 4.8.0 JuliaRegistries/General#162781; the other two clean-main initialization errors remain in a dedicated reproduce/bisect investigation.

Documentation and all branch-relevant QA, owner-clean import, platform-load, and formatting checks are green. No unrelated downstream workaround is being folded into this PR.

[ChrisRackauckas-Claude]

The downstream ModelingToolkit Julia 1.10 initialization failures have now been split precisely:

• 2 collector-invalidation errors are fixed in Fix collect_vars! after late invalidation on Julia 1.10 ModelingToolkit.jl#4852. The regression is bisected, covered on Julia 1.10/1.12, and the official Julia 1.10 InterfaceI group passes locally (1498 passed, 5 existing broken).
• the remaining 9 errors are the independent DelayDiffEq DEOptions resolver incompatibility tracked by Retroactively cap DelayDiffEq at OrdinaryDiffEqCore 4.8.0 JuliaRegistries/General#162781.

#4852 is a draft pending CI/review; this link is classification, not a claim that this PR's full matrix is green yet.