We use sentence splitting in the biaffine parser to keep the O(n^2)
biaffine attention model tractable. However, since the sentence splitter
makes errors, the parser may not have the correct head available.

This change adds another splitting strategy. The goal of this strategy
is to use the highest-probability splits to partition a doc until each
partition is smaller than or equal to a maximum length. This reduces
the number of attachment errors as a result of incorrect sentence
splits, while providing an upper bound on complexity.

The algorithm works as follows:

* If the length |d| > max_length:
  - Find the highest-probability split in d according to senter.
  - Split d into d_1 and d_2 using the highest probability split.
  - Recursively apply this algorithm to d_1 and d_2.
* Otherwise: do nothing

We use a back-off when the first token is the best splitting point, to
avoid an infinite recursion. The back-off was simply to use the second
token, refine this to choose the second-most probable splitting point.

We now use the spaCy parser scorer.

Now that Thinc doesn't set the Tensor type globally anymore, we have to
make sure that Tensors are placed on the correct device.

Before this change, we'd use the senter pipe directly. However, this did
not work with the transformer model without modifications (because it
clears tensors after backprop). By using the functionality proposed in

explosion/spaCy#11002

we can use the activations that are stored by the senter pipe in `Doc`.

This measure was removed.

Also make evaluation targets depend on the corpus they use.

Suggested by @kadarakos