[CustomOp] Fix ext_tensor.cast failed bug

[chenwhql]

PR types

Bug fixes

PR changes

Others

Describe

The code in macro __NVCC__ need to put into .cu file and compiled by nvcc.

#if defined(__NVCC__) || defined(__HIPCC__)
    } else if (platform::is_gpu_place(in_.place())) {
      platform::Transform<platform::CUDADeviceContext> trans;
      auto *context = static_cast<const platform::CUDADeviceContext *>(ctx_);
      trans(*context, in_begin, in_end, out_begin,
            CastDataTypeFunctor<InType, OutType>());
      context->Wait();
#endif

Before this PR, if we call tensor.cast(x) and x.place is kGPU, it will throw error.

[zhwesky2010]

LGTM

[jim19930609]

LoL, so we have a file named "ext_tensor.cu" for Linux Symlink, and a hidden ".ext_tensor.cu"(actually .ext_tensor.cu.cu) for Windows Symlink. What if there were a third/fourth... platform?

[zhwesky2010]

src is ext_tensor here. If system is based on Unix, these is no problem generally.

[chenwhql]

we should remove these symlink cases later, include data_type_transform, tensor_utils, etc.