Interleaved Multi-head Attention Operators (awslabs#884)

Replaced batched dot product in multi-head attention with interleaved_matmul attention operators to improve performance. Also changes the batch-major data to time-major format while in the model to comply with the new operator requirements.

Add ability to cache encoder outputs of model (awslabs#858)

Add SSRU layer and decoder (awslabs#851)

- Adds layers.SSRU, which implements a Simpler Simple Recurrent Unit as described by Kim et al, 2019.
- Adds ssru_transformer option to --decoder, which enables the usage of SSRUs as a replacement for the decoder-side self-attention layers.
- Reduces the number of arguments for MultiHeadSelfAttention.hybrid_forward().
previous_keys and previous_values should now be input together as previous_states, a list containing two symbols.

Fix batch sizing errors introduced in c00da52 (awslabs#849)

Option to suppress console output for secondary workers (awslabs#841)

Cross-entropy loss without SoftmaxOutput (awslabs#835)

Merge pull request awslabs#822 from awslabs/sockeye_2_merge_again

Exception merge commit to move master to Sockeye 2

Create sockeye.bib (awslabs#809)

An optional time limit for training in seconds (awslabs#715)

* Optional time limit for training.

* Changelog.

Update to pyyaml 5.1 (awslabs#690)

* update to pyyaml 5.1

* mypy