add TORCH_CUDA_CPP_API for AutoNcclGroup (pytorch#130012)

`torch::cuda::nccl` is an option for developers to depend only on torch but not nccl. But to use `torch::cuda::nccl::send`/`torch::cuda::nccl::recv`, `ncclGroupStart()`/`ncclGroupEnd()` is needed,  `torch::cuda::nccl::AutoNcclGroup` can be used.  but `torch::cuda::nccl::AutoNcclGroup` is not exported and is LOCAL symbol, which can't be used from outside of libtorch.

<img width="1618" alt="image" src="https://github.com/pytorch/pytorch/assets/1913192/25b0bd54-2da6-480f-876d-b05acfecfe62">

Pull Request resolved: pytorch#130012
Approved by: https://github.com/kwen2501, https://github.com/eqy

Author: houqi

torch/csrc/cuda/nccl.h

@@ -75,7 +75,7 @@ enum class ncclDataType {
 // RAII helper class to manage NCCL group API and CUDA free mutex.
 // The destructor is allowed to throw since this helper class only
 // manages group and lock lifetimes.
-struct AutoNcclGroup {
+struct TORCH_CUDA_CPP_API AutoNcclGroup {
   AutoNcclGroup();
   AutoNcclGroup(ncclComm_t comm, bool comm_nonblocking);
   ~AutoNcclGroup() noexcept(false);