Better docs of returned tensor in ctc_ops.py (#3062)

* Better docs of returned tensor in ctc_ops.py

'Logits' aren't a meaningful word for what the ctc_cost function returns. Browsing the implementation (tensorflow/tensorflow/core/util/ctc/) I saw that the cost function is returning the minus log probabilities of the target labelling, so, this new comment erases any doubt. Thanks!

* Update ctc_ops.py

tensorflow/python/ops/ctc_ops.py

@@ -95,7 +95,7 @@ def ctc_loss(inputs, labels, sequence_length,
     ctc_merge_repeated: Boolean.  Default: True.
 
   Returns:
-    A 1-D `float` `Tensor`, size `[batch]`, containing logits.
+    A 1-D `float` `Tensor`, size `[batch]`, containing the negative log probabilities.
 
   Raises:
     TypeError: if labels is not a `SparseTensor`.