Fix tf.keras.estimator.model_to_estimator in tf_nightly.

With the keras implementation moved the and the old estimator version here, model_to_estimator was reading the class definitions from the tensorflow source instead of the keras source, so some `isinstance()` checks were comparing two different implementations of the same class and returning False when it should have been True.

PiperOrigin-RevId: 380254729
Change-Id: If506ccdac39e9aacd9bc38b14608aaaf287e98a4

tensorflow/tools/pip_package/setup.py

@@ -119,7 +119,7 @@
     if 'tensorboard' in pkg:
       REQUIRED_PACKAGES[i] = 'tb-nightly ~= 2.6.0.a'
     elif 'tensorflow_estimator' in pkg:
-      REQUIRED_PACKAGES[i] = 'tf-estimator-nightly ~= 2.5.0.dev'
+      REQUIRED_PACKAGES[i] = 'tf-estimator-nightly ~= 2.6.0.dev'
     elif 'keras' in pkg and 'keras_preprocessing' not in pkg:
       REQUIRED_PACKAGES[i] = 'keras-nightly ~= 2.6.0.dev'