Revert ntree limit fix (#6616)

The old (before fix) best_ntree_limit ignores the num_class parameters, which is incorrect. In before we workarounded it in c++ layer to avoid possible breaking changes on other language bindings. But the Python interpretation stayed incorrect. The PR fixed that in Python to consider num_class, but didn't remove the old workaround, so tree calculation in predictor is incorrect, see PredictBatch in CPUPredictor.

tests/python/test_with_dask.py

@@ -933,9 +933,9 @@ class TestWithDask:
     def test_feature_weights(self, client: "Client") -> None:
         kRows = 1024
         kCols = 64
-
-        X = da.random.random((kRows, kCols), chunks=(32, -1))
-        y = da.random.random(kRows, chunks=32)
+        rng = da.random.RandomState(1994)
+        X = rng.random_sample((kRows, kCols), chunks=(32, -1))
+        y = rng.random_sample(kRows, chunks=32)
 
         fw = np.ones(shape=(kCols,))
         for i in range(kCols):