Do not use an explicit loop in unflatten

[danieldk]

...to avoid many small GPU <-> CPU transfers.

I still have to benchmark this change to see if this leads to a meaningful training time difference.

[danieldk]

@explosion-bot please test_gpu

[explosion-bot]

🪁 Successfully triggered build on Buildkite

URL: https://buildkite.com/explosion-ai/thinc-gpu-test-suite/builds/15

[svlandeg]

@danieldk : looks good! Feel free to merge in once you're done with your benchmarks and everything seems OK :-)

[danieldk]

Just a status update: I want to benchmark this a little more.

[danieldk]

I did more benchmarks and this change seems to make training slightly slower (2-3%) and inference slightly faster by the same amount. Since a declarative style is nicer in general and since this should scale better (the looping over the array will get more expensive with longer lengths arrays), I'll merge this.