Fix KeyedVectors.add_vectors() error when use most_similar

[s2terminal]

Fixing one of the issues in #3224

If I call most_similar() before doing add_vectors() and then call most_similar() again after doing add_vectors(), I get a ValueError: operands could not be broadcast together with shapes.
This error occurs because len(vectors) and len(vectors.norms) do not match.

from gensim.models import Word2Vec
import numpy

model = Word2Vec(sentences=[
                            ["this", "is", "test1"],
                            ["that", "is", "test2"],
], vector_size=2, min_count=1)

print(model.wv.most_similar("test1", topn=1)) #=> [('test2', 0.9941185712814331)]

model.wv.add_vectors(["test3"], [numpy.array([0.5, 0.5])])

print(model.wv.most_similar("test1", topn=1)) #=> ValueError: operands could not be broadcast together with shapes (6,) (5,) 

To resolve this error, I have used fill_norms to match len(vectors) and len(vectors.norms).

[mpenkov]

Needs a test. The example from the issue description is probably good enough.