Add constrained decoding to CRF

[napsternxg]

Describe the feature and the current behavior/state.
We can reuse code from the crf_multitag_sequence_score to allow constrained decoding.

Relevant information

• Are you willing to contribute it (yes/no): yes
• Are you willing to maintain it going forward? (yes/no): yes
• Is there a relevant academic paper? (if so, where): we only need to reuse the functions in tfa.text.crf to get this functionality
• Is there already an implementation in another framework? (if so, where): yes easy to extend tfa.text.crf
• Was it part of tf.contrib? (if so, where):

Which API type would this fall under (layer, metric, optimizer, etc.)
This will be under tfa.text.crf

Who will benefit with this feature?
People in NLP community, constrained decoding is important to deal with padding issues as well to denote start and end of a sequence.

Any other info.