Preserve Tracker gradients in GPUArraysCore restructure

[ChrisRackauckas-Claude]

Ignore this PR until it has been reviewed by @ChrisRackauckas.

Fixes #504

What changed and why

ArrayInterface.restructure on a GPUArraysCore.AbstractGPUArray target currently does Adapt.adapt(parameterless_type(x), y) then reshape. Tracker's adapt_structure is param(adapt(T, data(xs))), which starts a new tape leaf. The forward result stays numerically correct and still looks tracked, but the source gradient is silently zero.

This adds a GPUArraysCore+Tracker package extension that reshapes the TrackedArray instead, matching the existing Tracker methods for Array and Tracker.TrackedArray targets from #498. Untracked sources still go through Adapt so CPU→GPU conversion is unchanged.

Failing before / passing after

The new JLArray test (no GPU hardware) on unfixed restructure:

restructure GPUArraysCore + Tracker: Test Failed at test/ad.jl:69
  Expression: Array(Tracker.data(dx)) == ones(Float32, 6)
   Evaluated: Float32[0.0, 0.0, 0.0, 0.0, 0.0, 0.0] == Float32[1.0, 1.0, 1.0, 1.0, 1.0, 1.0]

Test Summary:                       | Pass  Fail  Total  Time
restructure GPUArraysCore + Tracker |    4     1      5  7.8s

With the extension:

Test Summary:                       | Pass  Total  Time
restructure GPUArraysCore + Tracker |    5      5  4.7s

Verification

$ GROUP=Core timeout 3600 julia --project=. -e 'using Pkg; Pkg.test()'
Test Summary:       | Pass  Total  Time
BandedMatrices      |   21     21  3.7s
BlockBandedMatrices |    8      8  5.9s
Core                |  221    221 27.8s
AD Integration      |   33     33 17.2s
StaticArrays        |   39     39  8.5s
ChainRules          |   20     20 22.9s
FillArrays          |    7      7  0.0s
Testing ArrayInterface tests passed

JuliaFormatter SciML style and typos were clean on the diff. git diff --check was clean.

Not verified locally

• GPU CI group (CUDA/Metal). This host has no NVIDIA device; the JLArray path is the device-independent stand-in from the issue.
• Downstream SciMLSensitivity TrackerAdjoint GPU job. The workaround in Preserve Tracker array parameter gradients SciML/SciMLSensitivity.jl#1636 should become unnecessary once this is registered, but that was not re-run here.
• Docs build: no docstring or public API change.
• ReverseDiff+GPU. ReverseDiff rarely wraps AbstractGPUArray; not in scope.

Reviewer notes

• The new method reshapes without Adapt, so a CPU TrackedArray restructured onto a GPU template stays a CPU tracked array. That matches restructure(::Array, ::TrackedArray) and is what preserves the tape. Untracked y still Adapts onto the GPU type.
• Test extra JLArrays is MIT (from GPUArrays.jl).
• Patch bump 7.30.0 → 7.30.1.

Links

• Issue: restructure drops Tracker gradients for GPUArraysCore targets #504
• Prior CPU Tracker/ReverseDiff restructure methods: Preserve tracked array wrappers in restructure #498
• Downstream TrackerAdjoint workaround: Preserve Tracker array parameter gradients SciML/SciMLSensitivity.jl#1636

🤖 Generated with Grok CLI 1.0.13 (model: grok-4.6)
Session: 01a05298-11d9-7a71-89a3-e977fe6a9a3e (local Grok CLI session; no shareable URL)

[codecov]

Codecov Report

❌ Patch coverage is 66.66667% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 61.35%. Comparing base (995c24d) to head (a5bc725).
⚠️ Report is 3 commits behind head on master.

Files with missing lines	Patch %	Lines
ext/ArrayInterfaceGPUArraysCoreTrackerExt.jl	66.66%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #505      +/-   ##
==========================================
+ Coverage   60.93%   61.35%   +0.42%     
==========================================
  Files          15       16       +1     
  Lines         622      634      +12     
==========================================
+ Hits          379      389      +10     
- Misses        243      245       +2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ChrisRackauckas-Claude]

The test (alldeps, 1.10, Core) Downgrade failure is pre-existing on master, not introduced here.

It is UndefVarError: OneElement not defined while precompiling ArrayInterfaceChainRulesExt (using ChainRules: OneElement) under lowest-compat deps. The same Downgrade workflow has been red on every recent master push, including https://github.com/JuliaArrays/ArrayInterface.jl/actions/runs/32094022221 (merge of #503).

A separate investigation is running in parallel to bisect/fix or file that master failure.

[ChrisRackauckas-Claude]

Follow-up: a reshape-only path kept the tape but left a CPU TrackedArray on the host. restructure now Adapts onto the GPU target type through a tracked primitive when the data is not already that type, and still skips Adapt.adapt on the TrackedArray itself (that goes through param() and zeros the source gradient).

CPU source onto a JLArray target now returns JLArray data with an all-ones gradient; GPU source still reshapes in place.