Write ELLPACK pages to disk (#4879)

* add ellpack source
* add batch param
* extract function to parse cache info
* construct ellpack info separately
* push batch to ellpack page
* write ellpack page.
* make sparse page source reusable

src/common/device_helpers.cuh

@@ -540,16 +540,21 @@ class BulkAllocator {
   }
 
  public:
-   BulkAllocator() = default;
+  BulkAllocator() = default;
   // prevent accidental copying, moving or assignment of this object
   BulkAllocator(const BulkAllocator&) = delete;
   BulkAllocator(BulkAllocator&&) = delete;
   void operator=(const BulkAllocator&) = delete;
   void operator=(BulkAllocator&&) = delete;
 
-  ~BulkAllocator() {
-    for (size_t i = 0; i < d_ptr_.size(); i++) {
-      if (!(d_ptr_[i] == nullptr)) {
+  /*!
+   * \brief Clear the bulk allocator.
+   *
+   * This frees the GPU memory managed by this allocator.
+   */
+  void Clear() {
+    for (size_t i = 0; i < d_ptr_.size(); i++) { // NOLINT(modernize-loop-convert)
+      if (d_ptr_[i] != nullptr) {
         safe_cuda(cudaSetDevice(device_idx_[i]));
         XGBDeviceAllocator<char> allocator;
         allocator.deallocate(thrust::device_ptr<char>(d_ptr_[i]), size_[i]);
@@ -558,6 +563,10 @@ class BulkAllocator {
     }
   }
 
+  ~BulkAllocator() {
+    Clear();
+  }
+
   // returns sum of bytes for all allocations
   size_t Size() {
     return std::accumulate(size_.begin(), size_.end(), static_cast<size_t>(0));