Improve/prune docs/tutorial of TranslationMatrix functionality

[gojomo]

The concerning test failure at #2944 now seems to me to be a false alarm. With more testing across many seeds, it appears the extremely flimsy BackMappingTranslationTest.test_infer_vector() was only passing in the base case (float64 randoms downcast to float32s) due to a lucky seeding, and only failing in the changed case due to unlucky seeding of the slightly-different stream of (float32 from the start) random numbers.

I've disabled the flimsy test, and it's questionable whether the BackMappingTranslationMatrix should even exist. It's perhaps 10 lines of using (not specializing-via-subclass) the actual TranslationMatrix class, and over-specialized on Doc2Vec models – whereas the TranslationMatrix functionality could and should be general to any vector-set, requiring just a few lines to apply to word-vectors, doc-vectors, or others. (And, calling the translation/projection infer_vector is unnecessarily prone to confusion with the different 'inference' that's native to Doc2Vec.)

I still think the TranslationMatrix itself is an under-appreciated bit of functionality, and I even strongly suspect – subject to experimentation – it could be part of a recommended solution for evolving a model to include more words that's far more robust/theoretically-defensible/performant than the build_vocab(..., update=True) & then incrementally .train() approach.

But, it'll need at the very least better docs/tutorial examples. The existing docs/notebook/tranlsation_matrix.ipynb is muddled & hard to run. (The test data it's using links to an all-in-Chinese Baidu download page that seems to require a login before raw .txt download.) It demos the BackmappingTranslationMatrix class in a later 'experimental' area I have trouble following even though it reuses some of the IMDB-dataset Doc2Vec tutorial I wrote.

I only have time to disable the BackMappingTranslationTest.test_infer_vector test right now, and this is pretty fringe functionality, so there's no urgency to clean it up - but this issue it to keep it under consideration, when the right person comes along.