pub fn should_parallelize(n_rows: usize, n_cols: usize) -> bool
Prefer serial when fork-join overhead exceeds the matvec work. ~256k element-ops matches the previous prefer_serial_matvec gate.
prefer_serial_matvec