LatentDirichletAllocation's components aren't normalized

An LDA model's components_ isn't easily interpretable:

>>> X = np.abs(np.random.RandomState(42).randn(5, 4))
>>> from sklearn.decomposition import LatentDirichletAllocation
>>> lda = LatentDirichletAllocation(n_topics=3).fit(X)
>>> lda.components_
array([[ 0.91847978,  2.43240154,  2.14285975,  1.10793777],
       [ 2.01874248,  0.92082054,  3.05363433,  3.58789672],
       [ 0.49890124,  0.50614032,  0.47761483,  0.51191327]])
>>> lda.components_.sum(axis=1)
array([ 6.60167884,  9.58109407,  1.99456967])

Looks like it's just not normalized, since normalizing in-place doesn't change transform outputs:

>>> d_before = lda.transform(X)
>>> lda.components_ /= lda.components_.sum(axis=1)[:, np.newaxis]
>>> d_after = lda.transform(X)
>>> norm(d_before - d_after)

But I don't know if a consequent partial_fit will still work.

Follow-up to #6320.

Start with LatentDirichletAllocation and inspect how components_, transform, and partial_fit are implemented. Check whether normalization is an intended invariant and whether changing components_ affects subsequent partial_fit calls. Done should resolve the normalization behavior while preserving correct transform results and include verification for the partial-fit case.