chunkize index[query] for large queries

[piskvorky]

... the same way chunking of the index itself is done, in for doc in index: aka SimilarityABC.__iter__.

In other words, don't construct the #query_docs x #index_docs result matrix at once, but chunkize query_docs and yield the result rows iteratively, in SimilarityABC.__getitem__.

Test thoroughly, because some code may depend on __getitem__ returning a matrix, not a generator.