Fixed copy constructor for HostDeviceVectorImpl. (#3657)

- previously, vec_ in DeviceShard wasn't updated on copy; as a result,
  the shards continued to refer to the old HostDeviceVectorImpl object,
  which resulted in a dangling pointer once that object was deallocated

src/common/host_device_vector.cu

@@ -58,6 +58,19 @@ struct HostDeviceVectorImpl {
       perm_d_ = vec_->perm_h_.Complementary();
     }
 
+    void Init(HostDeviceVectorImpl<T>* vec, const DeviceShard& other) {
+      if (vec_ == nullptr) { vec_ = vec; }
+      CHECK_EQ(vec, vec_);
+      device_ = other.device_;
+      index_ = other.index_;
+      cached_size_ = other.cached_size_;
+      start_ = other.start_;
+      proper_size_ = other.proper_size_;
+      SetDevice();
+      data_.resize(other.data_.size());
+      perm_d_ = other.perm_d_;
+    }
+
     void ScatterFrom(const T* begin) {
       // TODO(canonizer): avoid full copy of host data
       LazySyncDevice(GPUAccess::kWrite);
@@ -166,7 +179,12 @@ struct HostDeviceVectorImpl {
   // required, as a new std::mutex has to be created
   HostDeviceVectorImpl(const HostDeviceVectorImpl<T>& other)
     : data_h_(other.data_h_), perm_h_(other.perm_h_), size_d_(other.size_d_),
-      distribution_(other.distribution_), mutex_(), shards_(other.shards_) {}
+      distribution_(other.distribution_), mutex_() {
+    shards_.resize(other.shards_.size());
+    dh::ExecuteIndexShards(&shards_, [&](int i, DeviceShard& shard) {
+        shard.Init(this, other.shards_[i]);
+      });
+  }
 
   // Init can be std::vector<T> or std::initializer_list<T>
   template <class Init>