fix unflatten padding when last element is empty

[svlandeg]

Attempt to fix spaCy issue explosion/spaCy#3880. The error in that issue is raised by ops.pyx on the line assert len(X) == 0 (https://github.com/explosion/thinc/blob/master/thinc/neural/ops.pyx#L138), and it happens when the tagger model is run with an empty doc as last element in a call to nlp.pipe().

I think this happens because the flatten function adds one additional bit of padding at the end (https://github.com/explosion/thinc/blob/master/thinc/neural/ops.pyx#L119) which is not stripped away when unflatten removes the padding only when length != 0 (https://github.com/explosion/thinc/blob/master/thinc/neural/ops.pyx#L136) for the last variable length in lengths (accessed outside of its for loop).

I don't have a proper way of testing this yet, so I'm definitely not sure about this.

[honnibal]

This makes a lot of sense! I've hit this error as well recently with empty docs. It makes perfect sense that the error arises when the doc is last, thanks. It should be easy to add a test for this in the unit tests.