Better abstraction for keeping track of which constraints have been evaluated

[dance858]

Before this PR the problem class kept track on which constraints had to be re-evaluated in each iteration. It is a better abstraction to move this responsibility to the constraints themself. In particular, this is good for the separate Python interface in the future.

Claude below:

Summary

The last of the three values_version follow-ups (after #114 and #115). problem_jacobian kept its own affine shortcut: a problem-global jacobian_called bool, set after the first call, reset wholesale by problem_update_params. This PR replaces it with the counter scheme from #113:

• problem gains a per-constraint constraint_jac_seen array (allocated and seeded deliberately stale in problem_init_jacobian, freed in free_problem); the jacobian_called field is deleted.
• problem_jacobian now calls eval_jacobian on every constraint — affine ones return instantly via the impl-skip from affine optimization #114 — and memcpys a constraint's block into the aggregated Jacobian only when that constraint's values_version moved. The offset walk is unchanged (layout is fixed at init), and the assert(nnz_offset == J->nnz) stays.
• The flag reset in problem_update_params is deleted: expr_set_needs_refresh re-arms the constraint evals, which bump their versions, so the copy re-fires automatically — per constraint, rather than for all constraints at once.

Behavior: affine constraints skip both eval work and the memcpy after their first call, exactly as before but per-constraint and automatic; non-affine constraints bump every eval and copy every call, as before.

One pre-existing note (unchanged by this PR): the memcpy source c->jacobian->x assumes the value-buffer layout matches the init-time to_csr layout — true for sparse_matrix and permuted_dense, but would be wrong if a constraint's top-level Jacobian were ever a stacked_pd (block-major storage).

Tests

New test_problem_jacobian_memcpy_skip: an affine parameterized constraint (a*x) plus a non-affine one (exp(x)). After the first problem_jacobian, the affine slice of prob->jacobian->x is poisoned and a re-eval proves it is not rewritten while the non-affine slice is; a problem_update_params then restores the slice with the new parameter value.

Verification

• All 426 tests pass (Debug and UBSan builds)
• Guard-malloc + leaks clean: MallocScribble=1 MallocGuardEdges=1 leaks --atExit — 0 leaks
• PROFILE_ONLY build compiles
• clang-format applied; grep confirms no jacobian_called reference remains