pub fn f32_to_f16(x: f32) -> u16
x rounded to binary16, to nearest with ties to even, subnormals kept, past the top infinite.
x