Deterministic GPU histogram. (#5361)

* Use pre-rounding based method to obtain reproducible floating point
  summation.
* GPU Hist for regression and classification are bit-by-bit reproducible.
* Add doc.
* Switch to thrust reduce for `node_sum_gradient`.

python-package/xgboost/core.py

@@ -1682,7 +1682,6 @@ class Booster(object):
         if importance_type == 'weight':
             # do a simpler tree dump to save time
             trees = self.get_dump(fmap, with_stats=False)
-
             fmap = {}
             for tree in trees:
                 for line in tree.split('\n'):

python-package/xgboost/plotting.py

@@ -68,7 +68,9 @@ def plot_importance(booster, ax=None, height=0.2,
         raise ValueError('tree must be Booster, XGBModel or dict instance')
 
     if not importance:
-        raise ValueError('Booster.get_score() results in empty')
+        raise ValueError(
+            'Booster.get_score() results in empty.  ' +
+            'This maybe caused by having all trees as decision dumps.')
 
     tuples = [(k, importance[k]) for k in importance]
     if max_num_features is not None: