GPU performance logging/improvements (#3945)

- Improved GPU performance logging

- Only use one execute shards function

- Revert performance regression on multi-GPU

- Use threads to launch NCCL AllReduce

tests/cpp/tree/test_gpu_hist.cu

@@ -375,6 +375,7 @@ TEST(GpuHist, ApplySplit) {
 
   hist_maker.info_ = &info;
   hist_maker.ApplySplit(candidate_entry, &tree);
+  hist_maker.UpdatePosition(candidate_entry, &tree);
 
   ASSERT_FALSE(tree[nid].IsLeaf());