Downgrade: Core extras, initdt parse fix, AMF Krylov ref (no floors)

[ChrisRackauckas-Claude]

This PR should be ignored until reviewed by @ChrisRackauckas.

Summary

Substantive fixes for Downgrade CI without package version floor bumps.

Change	Why
Core: SparseArrays in [extras]	Test target was weakdep-only; Pkg.test aborts under Downgrade
ode_initdt_tests.jl: missing @testset end	ParseError on InterfaceI after #4001/#4000
AMF tests: ROS34PW1a(linsolve=KrylovJL()) for dense operator-jac refs	SciMLOperators WOperator convert MethodError on AddedOperator/tensor-product graphs (upstream SciMLOperators#411); adjoint FD tol 0.01→0.015

Sibling PR (floors only)

#4009 — Enzyme 0.13 → 0.13.180.

Independent of this PR; either order is fine.

[ChrisRackauckas-Claude]

Split this into one PR per issue, since the three changes here fail independent CI jobs and are independently reviewable:

• Close the unterminated @testset in ode_initdt_tests.jl #4013 — the missing @testset end in ode_initdt_tests.jl. Verified: InterfaceI now passes on julia 1 / lts / pre and Runic / Runic Format Check goes green.
• Core: declare SparseArrays in [extras] so the test target survives weakdep promotion #4014 — SparseArrays in [extras] for OrdinaryDiffEqCore. Verified: downgrade-sublibraries / test (lib/OrdinaryDiffEqCore) now passes.

The third change here (AMF linsolve = KrylovJL() + the 0.01 → 0.015 adjoint tolerance) I did not carry over. Root cause is upstream and already has PRs: WOperator's constructor leaves _concrete_form typed as an AddedOperator when J is one, so convert(AbstractMatrix, W) cannot assign the materialized matrix back — SciML/SciMLOperators.jl#410 / #411. Minimal repro, no OrdinaryDiffEq involved:

using SciMLOperators, LinearAlgebra   # v1.24.4
J = MatrixOperator(rand(4,4)) + MatrixOperator(rand(4,4))
W = SciMLOperators.WOperator{true}(I, 1.0, J, rand(4))
convert(AbstractMatrix, W)
# MethodError: Cannot `convert` an object of type Matrix{Float64}
#              to an object of type SciMLOperators.AddedOperator{...}

Once that ships, AMF should need only a SciMLOperators floor bump rather than a test-side workaround — worth avoiding the tolerance loosening if possible.

Also found while going through master: .typos.toml declares hom twice (lines 72 and 149, from #4004 and #4000), which is a TOML duplicate-key error, so Spell Check with Typos has been aborting before checking anything. Fixed in #4016. And the QA lane's AutoDePSpecialize rendered-docs failure is #4015.