Add onnx_trace argument for learned embeddings (#492)

Summary:
Pull Request resolved: https://github.com/pytorch/fairseq/pull/492

This argument was missing so we cannot export Transformer if we use learned positional embeddings. See also https://github.com/pytorch/translate/pull/335

Reviewed By: jhcross

Differential Revision: D13984781

fbshipit-source-id: 2187377e952ff587e07237de312c5b68f7d68891