v0.6.2

Changelog:

• 998ba4f: Add language models from Baevski & Auli (2018)
• 4294c4f: Add mixture of experts code from Shen et al. (2019)
• 0049349: Add example for multilingual training
• 48d9afb: Speed improvements, including fused operators from apex
• 44d27e6: Add Tensorboard support
• d17fa85: Add Adadelta optimizer
• 9e1c880: Add FairseqEncoderModel
• b65c579: Add FairseqTask.inference_step to modularize generate.py
• 2ad1178: Add back --curriculum
• Misc bug fixes and other features