Adds a linear-chain CRF layer to tensorflow/contrib.

This CRF is factored into unary scores for every tag from the layer below and transition scores between tags for adjacent words. For any tagging task given word sequences, this can simply replace the softmax layer. Optimizing the negative of tf.contrib.crf.crf_log_likelihood will maximize the probability of the given tag sequence.

This module also contains a Viterbi decoder that can be used to extract the sequence with the maximum likelihood at test time outside of TensorFlow.
Change: 133993074