Fix downstream test sources for OrdinaryDiffEq 7

[ChrisRackauckas-Claude]

Note

Please ignore this PR until reviewed by @ChrisRackauckas.

Summary

Follow-up to #593. That PR fixed the resolver, but the downstream test sources still don't run because the OrdinaryDiffEq 6→7 transition broke several call sites that the old resolver was masking. This PR makes the actual safetestsets pass.

Changes

test/downstream/odesolve.jl:

• Rosenbrock23(autodiff = false) → Rosenbrock23(autodiff = AutoFiniteDiff()). OrdinaryDiffEq 7 rejects Bool for that kwarg and asks for an ADType.
• Add using OrdinaryDiffEqRosenbrock (Rodas5 is no longer re-exported from OrdinaryDiffEq).
• Add using RecursiveArrayToolsShorthandConstructors (needed for the AP[...] syntax introduced in 334bd95; runtests.jl Pkg.develops the package but never using-s it).
• SciMLBase.successful_retcode(sol) → bare successful_retcode(sol). successful_retcode is reachable through using OrdinaryDiffEq; SciMLBase as a bare name is not. Avoiding using SciMLBase keeps SciMLBase/ADTypes out of [deps] — adding them as direct deps would unresolvably conflict with RAT 4.3's transitive SciMLBase 3 requirement.

test/downstream/downstream_events.jl:

• Add using RecursiveArrayToolsShorthandConstructors (same AP[...] reason).

test/downstream/Project.toml:

• Add OrdinaryDiffEqRosenbrock = "1, 2" and RecursiveArrayToolsShorthandConstructors.

Local verification

After this branch:

Test Summary: | Pass  Broken  Total     Time
ODE Solve     |    4       1      5  1m34.7s
Test Summary: | Total  Time
Events        |     0  2.4s
Test Summary:          | Broken  Total  Time
Measurements and Units |      1      1  4.2s
Test Summary: | Pass  Total  Time
TrackerExt    |    1      1  0.5s

All four Downstream-group safetestsets that runtests.jl exercises run to completion. The Broken totals are pre-existing @test_broken markers, not new failures.

Out of scope

• test/downstream/symbol_indexing.jl (SymbolicIndexingInterface group) still has many failures from MTK 11 / SII indexing API changes — its own investigation.
• test/downstream/adjoints.jl remains commented out in runtests.jl per the pre-existing TODO.

Test plan

• ODE Solve Tests safetestset passes locally (4 pass, 1 pre-existing broken)
• Event Tests with ArrayPartition safetestset runs without error
• Measurements and Units safetestset passes (1 pre-existing broken)
• TrackerExt safetestset passes (1 pass)
• CI Tests - Downstream - Julia 1/lts/pre

[ChrisRackauckas-Claude]

Update: pushed an additional fix for test/downstream/symbol_indexing.jl.

MTK 11 reorders unknowns(ODESystem(D(x), D(y))) to [y, x] (was [x, y] under MTK 8/9), which broke the hardcoded propertynames in the test_tables_interface call at line 80. Switched to deriving the names from unknowns(lv) so the test is order-agnostic.

After this commit, DiffEqArray Indexing Tests goes from 0/358 → 357/358 assertions passing on Julia 1.11 locally.

The remaining 1-test error is at line 103 (sol[x .+ [y, 2y, 3y]] ≈ vcat.(getindex.((sol,), [x[1] + y, x[2] + 2y, x[3] + 3y])...)) — exercises symbolic indexing with a Symbolics.Arr{Num, 1}. The test passes when run standalone but errors only inside @safetestset's isolated module (the stacktrace shows a BoundsError-style broadcast shape mismatch via SymbolicIndexingInterface.GetStateIndex). Looks like a separate Symbolics/SII module-isolation interaction that needs its own investigation — not gating the bulk of the downstream tests, and I didn't want to @test_broken it since that would only hide a real regression.

[ChrisRackauckas-Claude]

Update: pushed commit `39c3708` with a RAT-side workaround for the remaining `sol[x, :]` failure.

Root cause

The error was `DimensionMismatch` inside `SymbolicIndexingInterface.GetStateIndex(::Timeseries, prob, i)` at `SII/src/state_indexing.jl:43`:
```julia
getindex.(state_values(prob, i), gsi.idx) # missing 1-tuple wrap
```
When `gsi.idx` is a scalar Int this works (broadcasts length-1 against the timeseries), but when the index is array-symbolic (e.g. `x(t)[1:3]`), `gsi.idx` is a `Vector{Int}` and the broadcast tries to zip length-N timeseries against length-K idx vector → error. The companion no-`i` overload at line 38 already wraps correctly (`(gsi.idx,)`).

Fix split

• Upstream proper fix: SciML/SymbolicIndexingInterface.jl#157 — mirror the 1-tuple wrap.
• Here: route `A[sym, :]` through `A[sym]` at the RAT dispatch site, since the no-args symbolic getindex returns the same `Vector{Vector{Float64}}` and avoids the broken SII codepath. Once SII Remove unreachable code #157 lands and releases, this RAT short-circuit can be removed.

Verified

```
Test Summary: | Pass Total Time
DiffEqArray Indexing Tests | 358 358 13m09.4s
```