Docs: mention that we support float reductions when -ffast-math is used.