nir: Optimize double-precision lower_round_even()