Rewrite sparse dmatrix using callbacks. (#7092)

- Reduce dependency on dmlc parsers and provide an interface for users to load data by themselves.
- Remove use of threaded iterator and IO queue.
- Remove `page_size`.
- Make sure the number of pages in memory is bounded.
- Make sure the cache can not be violated.
- Provide an interface for internal algorithms to process data asynchronously.

tests/cpp/helpers.cu

@@ -8,16 +8,16 @@ namespace xgboost {
 
 CudaArrayIterForTest::CudaArrayIterForTest(float sparsity, size_t rows,
                                            size_t cols, size_t batches)
-    : rows_{rows}, cols_{cols}, n_batches_{batches} {
-  XGProxyDMatrixCreate(&proxy_);
-  rng_.reset(new RandomDataGenerator{rows_, cols_, sparsity});
+    : ArrayIterForTest{sparsity, rows, cols, batches} {
   rng_->Device(0);
   std::tie(batches_, interface_) =
       rng_->GenerateArrayInterfaceBatch(&data_, n_batches_);
   this->Reset();
 }
 
-CudaArrayIterForTest::~CudaArrayIterForTest() { XGDMatrixFree(proxy_); }
+size_t constexpr CudaArrayIterForTest::kRows;
+size_t constexpr CudaArrayIterForTest::kCols;
+size_t constexpr CudaArrayIterForTest::kBatches;
 
 int CudaArrayIterForTest::Next() {
   if (iter_ == n_batches_) {
@@ -28,8 +28,6 @@ int CudaArrayIterForTest::Next() {
   return 1;
 }
 
-size_t constexpr CudaArrayIterForTest::kRows;
-size_t constexpr CudaArrayIterForTest::kCols;
 
 std::shared_ptr<DMatrix> RandomDataGenerator::GenerateDeviceDMatrix(bool with_label,
                                                                     bool float_label,