* cast to higher precision type to prevent fp16 diff problem

* fix bilinear backward and add more unittest case