fix(worker): Adds before_update hook

[delucca]

Hi!

I found this repository on spacy#8093 and I was trying to train a large model in multiple GPUs.

I've tried using the latest Spacy version (3.5) with spacy-transformers and it seems this repo doesn't support it since the training.before_update key in the config is quite new (and required).

So, I've added it and I was able to launch a new training session with that change.

[delucca]

Just a note, I'm still trying to run it. For some reason spacy-ray still only allocates memory on a single GPU. But, either way, that fix is required (and quite easy) in any case

[rmitsch]

Hi @delucca, thanks for your contribution!

FYI, spacy-ray hasn't been updated recently and it was never quite ready for production use even for spacy v3.0. There are still a number of major limitations and bugs, see #14 (comment).

Failing tests

[svlandeg]

There's other things we need to fix, including the CI tests, but this PR by itself makes sense so I'll merge that in. Thanks!