fix: Ensure seed is based on RNG State

[KennethEnevoldsen]

The global seed (see also #942) cause variance in performance estimation depending on the order in which tasks are run, we see an example here:

embeddings-benchmark/results#19 (comment)

This PR solved the specific issue presented in the PR. In #942 I propose a more general fix to this issue, but this PR as well as using a consistent seed will naturally cause slightly different scores as RNG states change.

We might want to avoid merging this change until we change the major version (and then do #942 fully).

Checklist

• Run tests locally to make sure nothing is broken using make test.
• Run the formatter to format the code using make lint.

[KennethEnevoldsen]

@Samoed and @artemsnegirev would love your opinion on this as well

[Samoed]

Maybe it's better to use fix_seed from transformers, because it will fix seed for torch with different devices?

UPD. I read #942, and I think that part of fix_seed can be reused for fixing seed, because it use old format for numpy there

[KennethEnevoldsen]

Maybe it's better to use fix_seed from transformers, because it will fix seed for torch with different devices?

But this will still be influenced by the order of operations? (we essentially do this already within the abstasks). It will also be influenced by code-changes.