Switch TensorFlow's default checkpoint format to V2.

V2 delivers (1) up to 3x peak memory reduction during restore, (2) much lower
restore latency.  These have proven to be critical for production use cases.
Change: 137715407

tensorflow/python/training/saver.py

@@ -899,7 +899,7 @@ class Saver(object):
                builder=None,
                defer_build=False,
                allow_empty=False,
-               write_version=saver_pb2.SaverDef.V1,
+               write_version=saver_pb2.SaverDef.V2,
                pad_step_number=False):
     """Creates a `Saver`.