math: optimize lrint on 32bit targets