Fix external memory for get column batches. (#4622)

* Fix external memory for get column batches.

This fixes two bugs:

* Use PushCSC for get column batches.
* Don't remove the created temporary directory before finishing test.

* Check all pages.

include/xgboost/data.h

@@ -265,17 +265,7 @@ class SparsePage {
    * \brief Push one instance into page
    *  \param inst an instance row
    */
-  inline void Push(const Inst &inst) {
-    auto& data_vec = data.HostVector();
-    auto& offset_vec = offset.HostVector();
-    offset_vec.push_back(offset_vec.back() + inst.size());
-    size_t begin = data_vec.size();
-    data_vec.resize(begin + inst.size());
-    if (inst.size() != 0) {
-      std::memcpy(dmlc::BeginPtr(data_vec) + begin, inst.data(),
-                  sizeof(Entry) * inst.size());
-    }
-  }
+  void Push(const Inst &inst);
 
   size_t Size() { return offset.Size() - 1; }
 };