Inference support for Neural Shuffle-Exchange network

[EmilsOzolins]

When target dimensions are not known, use hparams.max_length. max_length should be a power of 2.
Max length can be enforced using hparams.force_max_length=True.
Auto-Regressive decoding not supported at this moment.

[dfurrer]

I get the following linter error (and a crash when trying to run). Could you fix this in the request please?

third_party/py/tensor2tensor/models/research/shuffle_network.py: 364: Type annotations not enabled. (<string>, line 364) [syntax-error]

[EmilsOzolins]

I hope this solves the problem.

[dfurrer]

Thanks. I have it running and the decode calls seem to work. However the linter now complains about these:
347: There should not be a space before a one-line docstring summary [g-space-before-docstring-summary]
347: One-line docstring summary should be followed by a blank line [g-no-space-after-docstring-summary]
389: There should not be a space before a one-line docstring summary [g-space-before-docstring-summary]
389: One-line docstring summary should be followed by a blank line [g-no-space-after-docstring-summary]
389: Missing Returns section in docstring, or wrong choice of Returns/Yields [g-doc-return-or-yield]

[EmilsOzolins]

Fixed those errors as well.

[afrozenator]

Many thanks to @OptimatiumFeles and @dfurrer both!