Fix Histogram allocation.

nidx_map is cleared after `Reset`, but histogram data size isn't changed hence
histogram recycling is used in later iterations.  After a reset(building new
tree), newly allocated node will start from 0, while recycling always choose
the node with smallest index, which happens to be our newly allocated node 0.

src/tree/updater_gpu_hist.cu

@@ -405,11 +405,18 @@ class DeviceHistogram {
     dh::safe_cuda(cudaSetDevice(device_id_));
     if (data_.size() >= kStopGrowingSize) {
       // Recycle histogram memory
-      std::pair<int, size_t> old_entry = *nidx_map_.begin();
-      nidx_map_.erase(old_entry.first);
-      dh::safe_cuda(cudaMemsetAsync(data_.data().get() + old_entry.second, 0,
-                                    n_bins_ * sizeof(GradientSumT)));
-      nidx_map_[nidx] = old_entry.second;
+      if (current_size + n_bins_ * sizeof(GradientSumT) < kStopGrowingSize) {
+        // just insert the new node.
+        nidx_map_[nidx] = current_size;
+        dh::safe_cuda(cudaMemsetAsync(data_.data().get() + current_size, 0,
+                                      n_bins_ * sizeof(GradientSumT)));
+      } else {
+        std::pair<int, size_t> old_entry = *nidx_map_.begin();
+        nidx_map_.erase(old_entry.first);
+        dh::safe_cuda(cudaMemsetAsync(data_.data().get() + old_entry.second, 0,
+                                      n_bins_ * sizeof(GradientSumT)));
+        nidx_map_[nidx] = old_entry.second;
+      }
     } else {
       // Append new node histogram
       nidx_map_[nidx] = current_size;