Fix lda transform prefomance

[menshikh-iv]

I founded one terrible performance issue connected with apply LDA model

Typical example

# *_PATH - path to file with model
# tokens - list with tokens

from gensim.corpora import Dictionary
from gensim.models import LdaModel

dct_model = Dictionary.load(DICT_PATH)
lda_model = LdaModel.load(LDA_PATH)

d2b_vector = dct_model.doc2bow(tokens)

print(lda_model[d2b_vector])

This triggers a chain of calls

__getitem__(self, bow, eps=None)

self.get_document_topics(bow, eps, self.minimum_phi_value, self.per_word_topics)

gamma, phis = self.inference([bow], collect_sstats=True)

collect_sstats=True initiates heavy computation for sstats, but sstats (phis) can't used if per_word_topics=False (Proof block)

I replaced True flag to per_word_topics for significant speedup get_document_topics for case per_word_topics=False (this effect is clearly visible if you use LdaModel with ~180 topics and 700k+ dictionary, i.e. huge topic matrix)

[tmylk]

Thanks for the fix. The docstrings in __init__ and get_item need to be updated to include the per_word_topics parameter.