push rope back to encoder/decoder

[vince62s]

@francoishernandez can you have a look for two things:

• not sure why I had to change again the random lm output of tests
• I did not touch the model config stuff, unsure if I have to.

EDIT:
ok the test failed because of the lm output. I think it comes from my pytorch version (I used nightly)

[francoishernandez]

Note for reference: the config already has the required information at encoder/decoder level, so no changes are needed. The issue in #206 was that the "model-level" rope was grabbing the "model-level" configuration, which can be misaligned with the nested encoder/decoder configurations if the latter are defined explicitly. Pushing rope back at encoder/decoder level means we always take the up to date config (either deduced from model-level, and forwarded in validation, or explicitly set).