Use SciMLSensitivity automatic sensealg choice

[ChrisRackauckas-Claude]

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

What changed and why

DeepEquilibriumNetworks no longer overrides SciMLSensitivity's automatic sensealg choice. solve is called with sensealg = nothing; a user-supplied sensealg in the layer constructor kwargs still wins. SciMLSensitivity is still imported so the solve adjoint rrules load.

For the out-of-place SteadyStateProblems this package constructs, that automatic choice is SteadyStateAdjoint(autodiff = false, autojacvec = ZygoteVJP()). The previous override kept autodiff = true on CPU (constructor default) and truncated the adjoint linear solve (maxiters = 10, 1e-3 tols). GPUArraysCore returns to a test extra; it was only needed for the GPU autodiff dispatch in the removed override.

This supersedes the GPU-only autodiff flip in #243: SciMLSensitivity already sets autodiff = false for GPU and out-of-place problems.

Verification

$ GROUP=QA julia +1.10 --project=. -e 'using Pkg; Pkg.test()'
Test Summary: | Pass  Total     Time
QA            |   18     18  3m25.0s
Testing DeepEquilibriumNetworks tests passed

$ GROUP=Core julia +1.10 --project=. -e 'using Pkg; Pkg.test()'
Test Summary: | Pass  Total   Time
Utils Tests   |   13     13  51.4s
Test Summary: | Pass  Total      Time
Layers Tests  | 1538   1538  23m10.9s
Testing DeepEquilibriumNetworks tests passed

Runic on the changed .jl files and typos over the diff both succeeded with no findings.

Docs: makedocs completed doctest, template expansion, citations, and document checks. The subsequent linkcheck step failed on a pre-existing ColPrac GitHub URL with HTTP 429 (rate limit), not on the api.md change. Two arXiv HTTP→HTTPS 301 warnings are also pre-existing.

Not verified locally

No CUDA GPU on this host, so the GPU test group was not run. Allowed-to-fail / downstream jobs were not run.

Behavior change

On large states (length(u0) > 50) this is not a no-op: the old truncated Krylov kwargs produced a different adjoint than SciMLSensitivity's default linear solve. That is intentional; the truncated solve was a DEQ-specific cheap adjoint, not a better general choice.

🤖 Generated with Grok Build (model: grok-4.6)
Agent-Session: local session ID 01a045fe-9bed-73c2-8681-bb5d74884034