Created by: kexinzhao

As mentioned in #9866 (closed), we want to have a float16 transpiler so that users could use it to run inference in float16 mode on the fly.