Created by: kexinzhao

fix #8607 (closed) fix #8608 (closed)

This PR does the following:

1. Make float16 compatible with data_type_transform mechanism as a prerequisite for operator's fp16 kernel selection.
2. Add isnan and isinf check for float16 to pass the TensorContainsNAN/INF test.
3. Fix header file include in float16.h to address the GPU segfault encountered when running tensor_util_test.