Use nccl group calls to prevent from dead lock. (#4113)

* launch all reduce sequentially.
* Fix gpu_exact test memory leak.

src/common/device_helpers.cuh

@@ -852,7 +852,7 @@ class AllReducer {
 #ifdef XGBOOST_USE_NCCL
   std::vector<ncclComm_t> comms;
   std::vector<cudaStream_t> streams;
-  std::vector<int> device_ordinals;
+  std::vector<int> device_ordinals;  // device id from CUDA
 #endif
 
  public:

src/common/host_device_vector.cu

@@ -459,9 +459,8 @@ HostDeviceVector<T>& HostDeviceVector<T>::operator=
 
 template <typename T>
 HostDeviceVector<T>::~HostDeviceVector() {
-  HostDeviceVectorImpl<T>* tmp = impl_;
+  delete impl_;
   impl_ = nullptr;
-  delete tmp;
 }
 
 template <typename T>