Avoid half-precision includes and functionality when CUDA_NO_HALF is defined

[eyalroz]

Sometimes, the half type defined in CUDA library conflicts with the one defined in other libraries. The CUDA_NO_HALF preprocessor define allows users of the CUDA APIs to circumvent this. In our code, however, we rely on cuda_fp16.hpp and on the half type being defined. Let's only have that reliance when CUDA_NO_HALF is undefined.