problem: build the Lagrange Hessian with sum_n_matrices_*

[dance858]

Moves problem.c's Lagrange Hessian assembly onto the library's sum_n_matrices_alloc / sum_n_matrices_fill_zero / sum_n_matrices_fill_add. The library is unchanged and the pin stays at cba21ac.

Changes

• Removed: problem_lagrange_hess_fill_sparsity (about 90 lines), hess_idx_map and the accumulator loop.
• lagrange_hessian: stays a CSR_matrix *, now the CSR view of the new hess_sum. The view is not a copy, and free_problem frees hess_sum. The bindings, the COO path and the tests don't change.
• Shared subexpressions: each Hessian is added right after it is evaluated, so constraints that share a node no longer sum overwritten values. This is the same pattern as Fix add's Hessian when its children share a subexpression #159 (add) and hstack: form the Hessian with sum_n_matrices #161 (hstack). The new test test_problem_hessian_shared_subexpr uses constraints log(x) and exp(log(x)) with the same log(x) node. On the old code it fails with -4 where -1 is correct.
• Affine constraints: they already skipped their eval, and now they skip the add too, since their Hessian is zero.
• problem.c no longer uses accumulator, compose_csr_idx_map_for_spd or csr_kernels.h. That clears the way for a library PR making them private.

Performance

Old main vs this branch, both Release, 5 interleaved runs. The Hessian checksums are identical.

n_vars, constraints	init_hessian old → new	problem_hessian old → new
1000, 3000 (+10 affine)	32–42 ms → 30–32 ms	1515–1543 µs → 1415–1459 µs
10, 50000 (+10 affine)	7.2–7.5 ms → 4.3–4.4 ms	624–670 µs → 573–609 µs

Testing

• All 303 tests pass in Debug, Release and an SP_TRACK_MEMORY build.
• leaks --atExit reports 0 leaks.

🤖 Generated with Claude Code