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

src/common/hist_util.cu

@@ -358,7 +358,7 @@ struct GPUSketcher {
       });
 
     // compute sketches for each shard
-    dh::ExecuteShards(&shards_, [&](std::unique_ptr<DeviceShard>& shard) {
+    dh::ExecuteIndexShards(&shards_, [&](int idx, std::unique_ptr<DeviceShard>& shard) {
         shard->Init(batch, info);
         shard->Sketch(batch, info);
       });