Run training with empty DMatrix. (#4990)

This makes GPU Hist robust in distributed environment as some workers might not
be associated with any data in either training or evaluation.

* Disable rabit mock test for now: See #5012 .

* Disable dask-cudf test at prediction for now: See #5003

* Launch dask job for all workers despite they might not have any data.
* Check 0 rows in elementwise evaluation metrics.

   Using AUC and AUC-PR still throws an error.  See #4663 for a robust fix.

* Add tests for edge cases.
* Add `LaunchKernel` wrapper handling zero sized grid.
* Move some parts of allreducer into a cu file.
* Don't validate feature names when the booster is empty.

* Sync number of columns in DMatrix.

  As num_feature is required to be the same across all workers in data split
  mode.

* Filtering in dask interface now by default syncs all booster that's not
empty, instead of using rank 0.

* Fix Jenkins' GPU tests.

* Install dask-cuda from source in Jenkins' test.

  Now all tests are actually running.

* Restore GPU Hist tree synchronization test.

* Check UUID of running devices.

  The check is only performed on CUDA version >= 10.x, as 9.x doesn't have UUID field.

* Fix CMake policy and project variables.

  Use xgboost_SOURCE_DIR uniformly, add policy for CMake >= 3.13.

* Fix copying data to CPU

* Fix race condition in cpu predictor.

* Fix duplicated DMatrix construction.

* Don't download extra nccl in CI script.

src/predictor/cpu_predictor.cc

@@ -60,6 +60,9 @@ class CPUPredictor : public Predictor {
       constexpr int kUnroll = 8;
       const auto nsize = static_cast<bst_omp_uint>(batch.Size());
       const bst_omp_uint rest = nsize % kUnroll;
+      // Pull to host before entering omp block, as this is not thread safe.
+      batch.data.HostVector();
+      batch.offset.HostVector();
 #pragma omp parallel for schedule(static)
       for (bst_omp_uint i = 0; i < nsize - rest; i += kUnroll) {
         const int tid = omp_get_thread_num();