Use the idf value from the largest index

[bratseth]

Today we take the idf value from the current index when needing it for relevance scoring of a document. With uncommon terms and small memory indexes this can lead to idf values which are far from those of a representative corpus.