Use Predictor for dart. (#6693)

* Use normal predictor for dart booster.
* Implement `inplace_predict` for dart.
* Enable `dart` for dask interface now that it's thread-safe.
* categorical data should be working out of box for dart now.

The implementation is not very efficient as it has to pull back the data and
apply weight for each tree, but still a significant improvement over previous
implementation as now we no longer binary search for each sample.

* Fix output prediction shape on dataframe.

tests/python/test_predict.py

@@ -22,6 +22,16 @@ def run_threaded_predict(X, rows, predict_func):
         assert f.result()
 
 
+def verify_leaf_output(leaf: np.ndarray, num_parallel_tree: int):
+    for i in range(leaf.shape[0]):     # n_samples
+        for j in range(leaf.shape[1]):  # n_rounds
+            for k in range(leaf.shape[2]):    # n_classes
+                tree_group = leaf[i, j, k, :]
+                assert tree_group.shape[0] == num_parallel_tree
+                # No sampling, all trees within forest are the same
+                assert np.all(tree_group == tree_group[0])
+
+
 def run_predict_leaf(predictor):
     rows = 100
     cols = 4
@@ -53,13 +63,7 @@ def run_predict_leaf(predictor):
     assert leaf.shape[2] == classes
     assert leaf.shape[3] == num_parallel_tree
 
-    for i in range(rows):
-        for j in range(num_boost_round):
-            for k in range(classes):
-                tree_group = leaf[i, j, k, :]
-                assert tree_group.shape[0] == num_parallel_tree
-                # No sampling, all trees within forest are the same
-                assert np.all(tree_group == tree_group[0])
+    verify_leaf_output(leaf, num_parallel_tree)
 
     ntree_limit = 2
     sliced = booster.predict(