Fix linear decay for learning rate in Doc2Vec.infer_vector. Fix #2061

[umangv]

This would fix issue #2061

[gojomo]

Looks good to me! This should be a high-priority fix. @menshikh-iv

[piskvorky]

Any updates needed to our tutorials / docs, after this change?

[gojomo]

This is always how it should have worked. If tests are passing (and they are), I'd expect any uses of inference in tutorials to just start working better – maybe a little, maybe a lot. (I wouldn't make merging this tiny fix dependent on refreshing all those result-cells.)

[menshikh-iv]

Thanks @umangv, congratz with the first contribution 🥇

[piskvorky]

@gojomo @umangv can you draft a summary of this PR for the release changelog? From the user perspective (not internal dev): who is affected, in what way, what should they do about it (if anything).

High-level; a few short, clear sentences.

[umangv]

This change fixes a bug in Doc2Vec.infer_vector that caused the learning rate to decay very fast. With this change, infer_vectors method should return more consistent document vectors and require far fewer steps to converge, especially for short documents.