cross_tower_ops -> cross_device_ops (#5776)

We've deprecated the "tower" terminology in DistributionStrategy, so the "cross_tower_ops" argument is now "cross_device_ops", matching the current name of "AllReduceCrossDeviceOps".

official/utils/misc/distribution_utils.py

@@ -42,7 +42,7 @@ def get_distribution_strategy(num_gpus, all_reduce_alg=None):
     if all_reduce_alg:
       return tf.contrib.distribute.MirroredStrategy(
           num_gpus=num_gpus,
-          cross_tower_ops=tf.contrib.distribute.AllReduceCrossDeviceOps(
+          cross_device_ops=tf.contrib.distribute.AllReduceCrossDeviceOps(
               all_reduce_alg, num_packs=2))
     else:
       return tf.contrib.distribute.MirroredStrategy(num_gpus=num_gpus)