Save average document length after fitting

[dunefox]

I think avgdl should be saved as an attribute after fitting so it's not estimated again if transform is called for one document instead of the 'training' corpus.

So

fit(X).transform(X)

makes sense because all documents in X use the same avgdl but

fit(X).transform(X)
transform(other_document)

then estimates avgdl for this document alone again.

[arosh]

You are completely right. It is terrible that this issue has been overlooked for more than several years...