This reverts commit 492f0cf9.

* Factor out validation and default opt getting
* Update test_models to use new parser.
* Remove unnecessary cast.
* Test translation server, format its docs.
* Fix Py27 compatibility.
* Actually fix Py27? And start testing TranslationServer.

* compute and store moving average in fp32

* In random sampling, make scores be score of EOS.

* Better buffering, readability, possible bug fix for return_attention
* Update docs.
* Test with return attention.
* Update docs, translation device selection, refactor translation strategies.

There is an unused argument in an example that seems to have been refactored out.

* Test length penalty.
* Fix #994 - now seems length penalty scores are consistent for fast and slow.
* Get matching summary cov score on fast.
* Get stepwise coverage penalty scores matching.
* Better document beam search.

* Remove non-base datasets.
* Update Dataset documentation.
* Move helper methods out of Dataset and document them.
* Remove , replace w explicit constructor calls.

* Experimental FP16 training
* Make batch size a multiple of 8 in case of FP16 training
* Force max_generator_batches to 0 in FP16 training
* Mark FP16 as experimental
* Make FP16 loss scale configurable
* Support FusedAdam from APEX
* Fix generator inspection index

* doc clean-ups, rename leagacy<>fast
* fix remove globalattn from transformer
* remove old beam calls -now fast is default
* keep old beam for coverage_penalty for now

* naive moving average on gpu
* validate with moving average if exists
* save moving average parameters instead of last update
* add average_every option

* Training option to iterate on the dataset once
* Add warning when single_pass is used but train_steps is non zero

In particular, the trainer should always save a checkpoint on the last
iteration. This PR also adds support for "-save_checkpoint_steps 0" to
only save the last checkpoint.

* Refactor, integrate penalties and global scorer into fast xlation.

* Update GlobalScorerStub for beam_search tests.

* Change min-length behavior, update tests, and document it.

* Simplify the training loop
* The outer `while steps <= train_steps` is not useful as the dataset
  yields an infinite iterator
* Multi-GPU sharding and batch accumulation can be expressed in terms
  of iterator wrappers
* Remove reduce_counter variable
* Rely on optim.training_step instead of manual increments

* Partial update to docs.

* Improve inputter documentation.

* Reverse FileNotFoundError instead of RuntimeError.

* add relative position representations for self attention.