[WIP] values_version cache coherence for Jacobian/Hessian value caches

[dance858]

Claude:
Replace the caller-must-refresh temporal contracts and the per-atom jacobian_csc_filled flags with a monotone values_version counter on matrix: writers bump, consumers refresh iff their recorded version differs. Staleness becomes impossible by construction and shared-operand double-refreshes dedupe for free.

• eval_jacobian/eval_wsum_hess free-function wrappers run the atom impl (slots renamed *_impl) and bump the output's version; the jacobian wrapper skips the bump for an affine node already evaluated this parameter epoch, preserving today's refresh counts.
• sparse_matrix csc_cache and stacked_pd csr_cache are version-guarded (the latter skips its per-call block memcpy in to_csr).
• spd vtable fill adapters and the raw hess_term2 write sites bump their outputs so version-guarded reads stay fresh.
• expr_set_needs_refresh gains a set_needs_refresh_children hook so the parameter-refresh walk reaches hstack's args[] children; without it a parameter-dependent affine child under hstack/vstack served stale spd Jacobian values after a parameter update (also fixes the pre-existing stale-CSC-mirror bug for args[] children on main).
• new tests lock the semantics: bump-per-eval, affine bump-skip and re-arm, CSC mirror dedup, spd to_csr freshness, spd hess-term staleness regression, and the hstack parameter-refresh regression.