Stack constraint and hstack Jacobians with vstack_matrices_*

[dance858]

Stacked on #164. Its base is bump-linalg-v0.5.0; retarget it to main once #164 merges.

Summary

This PR adopts the library's vstack_matrices_* (sparsediff_linalg #25) and finishes the left_matmul cleanup from #27.

• problem.c: the constraint Jacobian is now jac_stack = vstack_matrices_alloc(jac_ops, n_constraints), and prob->jacobian is its CSR view, so SparseDiffPy and jacobian_coo are unchanged. problem_jacobian evaluates the constraints and calls vstack_matrices_fill_values, which copies only the Jacobians whose values_version changed. This replaces constraint_jac_seen and both hand-written to_csr loops. With no constraints, jac_stack is an empty 0 × n_vars sparse matrix, because vstack needs at least one operand.
• hstack.c: the Jacobian is a vstack of the children's Jacobians (new field child_jac), so children whose values didn't change are no longer recopied on every eval.
• Fewer copies: vstack reads each operand's own storage. A stacked_pd Jacobian is now copied once per change instead of twice (once into its CSR view, then once into the stack), and no CSR cache is kept for it.
• left_matmul: removes Jchild_CSC, J_CSC, csc_to_csr_work and the kron-only workspace. All of them have been dead since the kron branch moved to BA_sparse_kron_matrices_*. subexpr.h no longer includes CSC_matrix.h.

Testing

• New test: test_jacobian_hstack_stacked_pd_child builds hstack(A @ exp(X), sin(X)). The left_matmul child's Jacobian is a stacked_pd, and the test checks the result against finite differences at two points, so the refill is exercised.
• Existing tests also cover the new paths:
  • test_problem_jacobian_memcpy_skip still passes. It writes into the affine slice directly and checks that it isn't recopied.
  • So does test_problem_jacobian_spd_constraint_interleaved.
• Builds: Debug and Release run 304 engine tests (303 plus the new one) and 201 library tests, all passing. PROFILE_ONLY=ON builds, the warnings are the same 5 as on main, and clang-format is clean.
• SparseDiffPy smoke test: I built SparseDiffPy against this branch in a throwaway venv. A problem with hstack, diag_vec and affine constraints gave a 28×4 Jacobian that matches finite differences at two points (max error 2e-10). Its own pytest suite can't run locally: the tests need a scope fixture that isn't defined anywhere in the repo.

🤖 Generated with Claude Code