Add HOWTO: Data-parallel training

[Marc van Zee]

See #1982, this HOWTO will implement data-parallel training using pmap in the seq2seq example.

[Marc van Zee]

@gmittal mentioned they would like to implement this based on #1982

[Marc van Zee]

@gmittal any updates on this?

[Marc van Zee]

FYI I am closing this issue in favor of #2116

[Gautam Mittal]

@marcvanzee sorry for delay -- this is still in-progress and will have a PR for this in the next week.