ensure consistency of nO dim for biLSTM

[svlandeg]

Fixes explosion/spaCy#7088

When using a biLSTM, the given nO dimension of the Thinc model would be divided by half and stored as such. I think this behaviour is unexpected, and it clashes with some of the shape inference & listener code in spaCy.

I think it makes more sense to keep the given nO, but provide the Torch model with half the value for its hidden features, so that the final output matches the required nO dimension.

This PR changes the behaviour both for PyTorchLSTM and LSTM (if bi is True) - both would otherwise predict vectors that were incompatible with their internal nO value - cf test and spaCy issue linked above.

The additional line in the test makes CauchySimilarity error as well. Not sure (yet) whether this is due to an error in get_width or the model itself, or whether this model should just be excluded from this particular line in the test. Suggest to address this in a follow-up PR.

[codecov]

Codecov Report

Merging #484 (a35ad58) into master (1d998ad) will decrease coverage by 0.36%.
The diff coverage is 88.46%.

@@            Coverage Diff             @@
##           master     #484      +/-   ##
==========================================
- Coverage   59.31%   58.94%   -0.37%     
==========================================
  Files         102      102              
  Lines        7206     7207       +1     
==========================================
- Hits         4274     4248      -26     
- Misses       2932     2959      +27     

Impacted Files	Coverage Δ
thinc/tests/layers/test_layers_api.py	0.00% <0.00%> (ø)
thinc/layers/lstm.py	98.18% <100.00%> (-0.02%)	⬇️
thinc/layers/cauchysimilarity.py	32.43% <0.00%> (-67.57%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1d998ad...a35ad58. Read the comment docs.

[honnibal]

Yeah this is sensible, I agree that a bunch of places expect nO to behave a certain way.