fix engine crash in shutdown phase

src/engine/threaded_engine.h

@@ -38,6 +38,7 @@
 #include <mutex>
 #include <string>
 #include <thread>
+#include <mxnet/storage.h>
 #include "./engine_impl.h"
 #include "../profiler/profiler.h"
 #include "./openmp.h"
@@ -306,6 +307,8 @@ class ThreadedEngine : public Engine {
  objpool_varblk_ref_ = common::ObjectPool<VersionedVarBlock>::_GetSharedRef();
  objpool_var_ref_ = common::ObjectPool<ThreadedVar>::_GetSharedRef();
 
+ storage_ref_ = Storage::_GetSharedRef();
+
  // Get a ref to the profiler so that it doesn't get killed before us
  profiler::Profiler::Get(&profiler_);
  }
@@ -549,6 +552,12 @@ class ThreadedEngine : public Engine {
  std::shared_ptr<common::ObjectPool<VersionedVarBlock> > objpool_varblk_ref_;
  std::shared_ptr<common::ObjectPool<ThreadedVar> > objpool_var_ref_;
 
+ /*!
+ * \brief Async destruction of some objects is relied on storage,
+ * prevent it from being destructed too early
+ */
+ std::shared_ptr<Storage> storage_ref_;
+
 #if MXNET_USE_CUDA
  /*! \brief Number of GPU devices available */
  std::atomic<int> device_count_{-1};