pub fn f16_bytes_to_f32(bytes: &[u8]) -> Vec<f32>
Decode a float16 byte slice into f32 values, accumulating in f32 for downstream dot-product accuracy.