Merging chironomid training sets: implications for palaeoclimate reconstructions