Use dlopen to load NCCL. (#9796)

This PR adds optional support for loading nccl with `dlopen` as an alternative of compile time linking. This is to address the size bloat issue with the PyPI binary release.
- Add CMake option to load `nccl` at runtime.
- Add an NCCL stub.

After this, `nccl` will be fetched from PyPI when using pip to install XGBoost, either by a user or by `pyproject.toml`. Others who want to link the nccl at compile time can continue to do so without any change.

At the moment, this is Linux only since we only support MNMG on Linux.

src/common/device_helpers.cuh

@@ -115,30 +115,6 @@ XGBOOST_DEV_INLINE T atomicAdd(T *addr, T v) {  // NOLINT
 }
 namespace dh {
 
-#ifdef XGBOOST_USE_NCCL
-#define safe_nccl(ans) ThrowOnNcclError((ans), __FILE__, __LINE__)
-
-inline ncclResult_t ThrowOnNcclError(ncclResult_t code, const char *file, int line) {
-  if (code != ncclSuccess) {
-    std::stringstream ss;
-    ss << "NCCL failure: " << ncclGetErrorString(code) << ".";
-    ss << " " << file << "(" << line << ")\n";
-    if (code == ncclUnhandledCudaError) {
-      // nccl usually preserves the last error so we can get more details.
-      auto err = cudaPeekAtLastError();
-      ss << "  CUDA error: " << thrust::system_error(err, thrust::cuda_category()).what() << "\n";
-    } else if (code == ncclSystemError) {
-      ss << "  This might be caused by a network configuration issue. Please consider specifying "
-            "the network interface for NCCL via environment variables listed in its reference: "
-            "`https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html`.\n";
-    }
-    LOG(FATAL) << ss.str();
-  }
-
-  return code;
-}
-#endif
-
 inline int32_t CudaGetPointerDevice(void const *ptr) {
   int32_t device = -1;
   cudaPointerAttributes attr;