[coz3-deepperf-fix] Batch fixed-column bound-witness linearization per row in lar_solver (#10029)

### Summary

`lp_bound_propagator::explain_fixed_in_row` explained every fixed column
of a row
independently, calling `lar_solver::explain_fixed_column` once per fixed
column
(`src/math/lp/lar_solver.cpp`). Each such call linearizes the lower- and
upper-bound witnesses of a single column — a BFS over the `u_dependency`
DAG using
the dependency manager's mark bits — and inserts every reached leaf
constraint into
the `explanation`.

Fixed columns of the same row routinely share large portions of their
bound-witness
sub-DAGs (common ancestor constraints). The per-column scheme therefore
re-traverses
those shared sub-DAGs and re-inserts their leaves once for *every*
column, with an
independent mark/unmark cycle per column.

### Change

Add `lar_solver::explain_fixed_in_row(row, ex)`, which collects the
lower/upper
witnesses of all fixed columns in the row and linearizes them together
in a single
`u_dependency_manager::linearize` pass.
`lp_bound_propagator::explain_fixed_in_row`
and `explain_fixed_in_row_and_get_base` now delegate to it; the
base-column lookup in
the latter is unchanged. `explain_fixed_column` is kept for its
single-column caller.

### Why it is correct

`explanation` is a set — `push_back` deduplicates. Dependency
reachability is
monotone, so the union of the per-column leaf sets equals the leaf set
of the union
of all roots: the batched pass yields exactly the same explanation. The
manager's
mark bits guarantee each shared sub-DAG node is visited once, and the
`linearize(ptr_vector, ...)` overload already skips null/duplicate
roots.

### Complexity

For a row with `N` fixed columns:

- before: `O(Σ_j |witness-DAG(j)|)` traversal + `O(Σ_j leaves(j))` set
insertions, with `N` mark/unmark cycles;
- after: `O(|⋃_j witness-DAG(j)|)` traversal + `O(#distinct leaves)` set
insertions, with a single mark/unmark cycle.

Shared sub-DAGs are walked and their leaves inserted once instead of
once per column.

### Measured effect

Profiled with callgrind on a representative conflict-heavy `QF_SLIA`
input
(`model_validate=true`, bounded run), baseline vs. patched:

- `lp::lar_solver::explain_fixed_column` on the hot path:
`24,337,671,616 → 0`
retired instructions (59.2% → 0% of the run), replaced by the single
batched
  traversal;
- total retired instructions: `41,113,093,210 → 35,983,363,256` (×0.875,
≈ 12.5%
  fewer) — the net work removed by de-duplicating shared sub-DAGs;
- wall-clock: `6.428 s → 6.079 s` (≈ 5.4% faster);
- differential correctness preserved (identical results across the
validation inputs).

<!-- gh-aw-workflow-id: coz3-deepperf-fix -->
<!-- gh-aw-workflow-call-id: Z3Prover/bench/coz3-deepperf-fix -->

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

src/math/lp/lp_settings.cpp

@@ -46,6 +46,7 @@ void lp::lp_settings::updt_params(params_ref const& _p) {
     m_dio_calls_period_decrease = lp_p.dio_calls_period_decrease();
     m_dio_run_gcd = lp_p.dio_run_gcd();
     m_random_hammers = lp_p.random_hammers();
+    m_batch_explain_fixed_in_row = lp_p.batch_explain_fixed_in_row();
     m_lcube = lp_p.lcube();
     m_lcube_flips = lp_p.lcube_flips();
     unsigned hammer_period = lp_p.int_hammer_period();