Optimize adapter element counting on GPU. (#9209)

- Implement a simple `IterSpan` for passing iterators with size.
- Use shared memory for column size counts.
- Use one thread for each sample in row count to reduce atomic operations.

tests/cpp/data/test_device_adapter.cu

@@ -51,3 +51,22 @@ void TestCudfAdapter()
 TEST(DeviceAdapter, CudfAdapter) {
   TestCudfAdapter();
 }
+
+namespace xgboost::data {
+TEST(DeviceAdapter, GetRowCounts) {
+  auto ctx = MakeCUDACtx(0);
+
+  for (bst_feature_t n_features : {1, 2, 4, 64, 128, 256}) {
+    HostDeviceVector<float> storage;
+    auto str_arr = RandomDataGenerator{8192, n_features, 0.0}
+                       .Device(ctx.gpu_id)
+                       .GenerateArrayInterface(&storage);
+    auto adapter = CupyAdapter{str_arr};
+    HostDeviceVector<bst_row_t> offset(adapter.NumRows() + 1, 0);
+    offset.SetDevice(ctx.gpu_id);
+    auto rstride = GetRowCounts(adapter.Value(), offset.DeviceSpan(), ctx.gpu_id,
+                                std::numeric_limits<float>::quiet_NaN());
+    ASSERT_EQ(rstride, n_features);
+  }
+}
+}  // namespace xgboost::data