[Feature Request] Support setting `training=False` argument in `BeamSearchDecoder`

[HanGuo97]

System information

• TensorFlow version (you are using): 2.0-beta
• TensorFlow Addons version: 0.40
• Is it in the tf.contrib (if so, where): No
• Are you willing to contribute it (yes/no): No
• Are you willing to maintain it going forward? (yes/no): No

Describe the feature and the current behavior/state.
Current BeamSearchDecoder does not support setting training=False argument during the computation. This limits many scenarios in which model behaves differently between training/inference stage, e.g., dropout.

Will this change the current api? How?
An extra argument that supports setting the training argument should be added to this line

Who will benefit with this feature?
Everyone who wants the RNN Cell to behave differently between training and inference (e.g., dropout).
Any Other info.

[seanpmorgan]

CC some people more familiar with seq2seq @qlzh727 @guillaumekln

[guillaumekln]

Yes, it seems a training flag should be forwarded from BaseDecoder.call to the cell call of concrete decoders (BasicDecoder and BeamSearchDecoder).