Fix Runic format check: add return in test/runtests.jl core_body()

[ChrisRackauckas-Claude]

Problem

Red checks on this branch / on main:

1. Runic Format Check — test/runtests.jl: the final expression of core_body() was missing the explicit return Runic requires.
2. GPU Tests — ArgumentError: Package LuxCUDA not found: LuxCUDA was dropped from [extras]/test [targets] of Project.toml during the v1.2 CI conversion (Canonical CI: grouped-tests.yml + root test/test_groups.toml #217).
3. Documentation — two genuine @example failures in the basic MNIST DEQ tutorial (reproduce on main, independent of the test/Project.toml changes here).

Fixes in this PR

Runic — ran Runic in-place on test/runtests.jl to add the return.

LuxCUDA test dep — restored LuxCUDA (compat 0.3) to [compat], [extras], and the test target (correct registry UUID d0bbae9a-e099-4d5b-a835-1c6931763bda).

Documentation — fixed the two real tutorial bugs:

• Steady-state adjoint DimensionMismatch. For a conv DEQ the rootfind state is a 4D array ((13, 13, 64, 2)). When the state has > 50 elements, SciMLSensitivity's SteadyStateAdjoint takes the matrix-free VecJacOperator path and builds the operator from vec(y), so it seeds the residual's pullback with a flat cotangent. Zygote's ProjectTo then rejects that flat cotangent when broadcast against the multi-dimensional u in the residual y .- u, raising variable with size(x) == (13,13,64,2) cannot have a gradient with size(dx) == (21632,). The Core suite never hit this (its conv states are ≤ 50 elements → dense-Jacobian path). Fix in src/layers.jl: compute the residual in flattened space, reshape(vec(y) .- vec(u), size(u)) — a value-preserving no-op that keeps the broadcast operands consistent with the flat cotangent.
• UndefVarError: VCAB3. The tutorial calls VCAB3() but only imported OrdinaryDiffEq, which (v7) no longer re-exports VCAB3. Added OrdinaryDiffEqAdamsBashforthMoulton to docs/Project.toml and using OrdinaryDiffEqAdamsBashforthMoulton: VCAB3 to the tutorial.

Bumped version to 2.7.1 (source behavior change).

Local verification (Julia 1.12, the failing 1 channel)

• Full docs/make.jl builds end-to-end: ExpandTemplates (both MNIST @example blocks, incl. train_model(NewtonRaphson(...)) and train_model(VCAB3(), :deq)) → RenderDocument → generated tutorials/basic_mnist_deq/index.html. No DimensionMismatch, no UndefVarError. (Only difference from CI: a transient ColPrac linkcheck 429 rate-limit from the sandbox; re-running with linkcheck=false builds cleanly, exit 0.)
• GROUP=Core Pkg.test() passes: Utils 11/11, Layers 1538/1538.
• Runic --check src test docs: exit 0.

Please ignore until reviewed by @ChrisRackauckas.