call will call build

Make the requirement clear. We don't have to build before __call__.

tensorflow/python/keras/engine/base_layer.py

@@ -924,6 +924,7 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
         for `inputs` by the previous layer (if `input` did come from
         a layer that generated a corresponding mask, i.e. if it came from
         a Keras layer with masking support.
+      - If the layer is not built, the method will call `build`.
 
     Raises:
       ValueError: if the layer's `call` method returns None (an invalid value).