Round to nearest integer instead of truncating in `f2fx` to ensure correct round-trip with `fx2f`. Add test to ensure correct round-trip with maximum `0.0005f` absolute error.
