test(reduction): use sol.u[end] for state-vector output

[ChrisRackauckas-Claude]

Note: Please ignore this PR until reviewed by @ChrisRackauckas. Opening as a draft.

Summary

Fixes the Reduction testset error reported on the arctic1-1
self-hosted runner:

Reduction: Error During Test at .../test/reduction.jl:48
  Test threw exception
  Expression: sum(sim1.u) ≈ sim2.u
  MethodError: no method matching isapprox(::Float64, ::Vector{Float64})

Root cause: under SciMLBase v3, last(sol) for a 1-component ODE
now returns a Float64 scalar rather than a 1-element
Vector{Float64}. With the existing output_func that flowed
through to:

• sim1.u :: Vector{Float64} (was Vector{Vector{Float64}})
• sum(sim1.u) :: Float64 (was Vector{Float64})
• sim2.u :: Vector{Float64} (1-element, from u_init = [0.0])

so @test sum(sim1.u) ≈ sim2.u ends up comparing a scalar to a
1-element vector and errors. The test file wasn't migrated for the
shape change when the rest of the suite moved to SciMLBase v3 /
RAT v4.

Switching output_func to return sol.u[end] (still a
Vector{Float64}) keeps each per-trajectory output as a state
vector, restoring the original shape semantics so the three
existing assertions hold without relaxing what they check.

Test plan

• GROUP=JLArrays julia +1.10 --project=test running the
  Reduction @safetestset:

  Test Summary: | Pass  Total   Time
  Reduction     |    3      3  24.1s
  

• Runic format-check passes (exit 0).
• Self-hosted CUDA runner re-runs Reduction.