Fix Tracker gradient through solve on RecursiveArrayTools v4

[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

Summary

Fixes the Tracker outer-AD gradient failures tracked in
SciMLSensitivity.jl#1331. The issue was filed as "Tracker tests fail on Julia 1.12+", but bisecting the actual failure shows the trigger is the RecursiveArrayTools major version bump, not Julia.

In RAT v4, AbstractVectorOfArray <: AbstractArray, so the generic sol isa AbstractArray branch in the Tracker @grad function DiffEqBase.solve_up now matches ODESolution and returns the nested sol.u — a Vector{Vector{Float64}}. Tracker tracks that vector-of-vectors directly, and downstream sum(solve(...)) reduces the outer vector element-wise into a Vector{Float64}, so Tracker.gradient(loss, p) trips its losscheck with "Function output is not scalar".

On RAT v3, AbstractVectorOfArray was not an AbstractArray, so the same code path fell through to convert(AbstractArray, sol), which RAT v3 implemented as stack(VA.u) — a flat Matrix{Float64}. That matrix tracked correctly and sum reduced to a scalar.

Fix

Detect AbstractVectorOfArray first in the Tracker @grad and Array(sol) (i.e. stack(sol.u)) it before handing the value to Tracker. This restores the pre-RAT-v4 behavior with no changes to the API.

if sol isa RecursiveArrayTools.AbstractVectorOfArray
    return Array(sol), pb_f
end

RecursiveArrayTools is already a direct [deps] of DiffEqBase (compat "4"), so no compat changes are needed.

Repro

On the current package set (Julia 1.12.6, Tracker 0.2.38, RAT 4.3.0, SciMLSensitivity master):

using SciMLSensitivity, OrdinaryDiffEq
import Tracker

function fiip(du, u, p, t)
    du[1] = p[1]*u[1] - p[2]*u[1]*u[2]
    du[2] = -p[3]*u[2] + p[4]*u[1]*u[2]
end
prob = ODEProblem(fiip, [1.0, 1.0], (0.0, 10.0), [1.5, 1.0, 3.0, 1.0])

loss = p -> sum(solve(prob, Tsit5(); p, saveat=0.1,
                      abstol=1e-10, reltol=1e-10,
                      sensealg=InterpolatingAdjoint()))

Tracker.gradient(loss, [1.5, 1.0, 3.0, 1.0])
# ERROR: Function output is not scalar

With this PR all five originally-failing tests in SciMLSensitivity's test/concrete_solve_derivatives.jl (save_idxs, save_everystep=false, non-integer saveat=2.3, VecOfArray, BouncingBall) pass with @test_broken false; continue guards removed.

Test plan

• Existing DiffEqBase tests pass.
• Follow-up PR in SciMLSensitivity.jl removes the @test_broken false Tracker guards in test/concrete_solve_derivatives.jl and the originally-failing tests pass.