gallivm: use nearest rounding for float->unorm24 conversion