[MRG] fix for loading old word2vec models

[jayantj]

Fixes loading Word2Vec models from previous versions.

While loading pre-keyedvector word2vec models, the attributes which belonged to the Word2Vec instance before the change and to the KeyedVector instance after the change, need to be set manually after unpickling the model. These were not being set correctly for models which saved the numpy data in separate files from the model file.
There was a unittest to verify this behaviour explicitly, but the existing model in the test data seems to contain a KeyedVector instance already.

I've updated the test data with a model without a keyedvector instance, and the code for loading models to handle this behaviour correctly.

Fixes #1173 (tested by loading the model linked in the issue and performing a lookup for "bon")