Created by: kexinzhao

Instead of using cast op, we check the expected kernel type (including input tensor data type) of an operator and the actual input tensor, if there is a mismatch on the data type, we do a data_type_transform on the input tensor (say transform the data type from fp32 to fp16) before we run the compute kernel of that operator.

We want to make float16 compatible with this data_type_transform mechanism.

There is a WIP pr working on this.