Make "Neg" op hidden from Python.

While this technically changes the API. This was in the TensorFlow 1.0 NumPy
design, so this is a bug fix to that plan, rather than a new API change.
Change: 144898546

tensorflow/python/ops/hidden_ops.txt

@@ -237,6 +237,7 @@ Max
 Mean
 Min
 Mul
+Neg
 Pow
 Prod
 Range

tensorflow/python/ops/math_ops.py

@@ -403,11 +403,11 @@ def negative(x, name=None):
   """
   with ops.name_scope(name, "Neg", [x]) as name:
     if isinstance(x, sparse_tensor.SparseTensor):
-      x_neg = gen_math_ops.neg(x.values, name=name)
+      x_neg = gen_math_ops._neg(x.values, name=name)
       return sparse_tensor.SparseTensor(
           indices=x.indices, values=x_neg, dense_shape=x.dense_shape)
     else:
-      return gen_math_ops.neg(x, name=name)
+      return gen_math_ops._neg(x, name=name)
 # pylint: enable=g-docstring-has-escape
 
 
@@ -861,7 +861,7 @@ def to_bfloat16(x, name="ToBFloat16"):
   return cast(x, dtypes.bfloat16, name=name)
 
 
-ops.Tensor._override_operator("__neg__", gen_math_ops.neg)
+ops.Tensor._override_operator("__neg__", gen_math_ops._neg)
 ops.Tensor._override_operator("__abs__", abs)
 # __invert__ corresponds to the ~ operator.  Here we follow the numpy convention
 # ~ marks an elementwise bit-wise inverse.  This is only implemented for boolean