time_major parameter for ctc_loss op (#3993)

* time_major parameter for ctc_loss op

Just as dynamic_rnn and to be more regular across the TensorFlow's code I added a flag capable of transpose the input data to be in time_major, required by gen_ctc_ops._ctc_loss. Note that, instead of set time_major=False (as the most of TensorFlow data), it was set False to be compatible with previous ctc_loss function call (but I don't think that this is a good idea).

* Fixing typos in docstring