Remove expensive prediction run during test collection (#279)
Removes the expensive computation during test collection. Such a collection is very problematic because:
1) Makes refreshing tests incredibly long (10+mins)
2) The collection should be light, so heavy computations there are very unexpected and can cause undisired behavior (e.g. refreshing tests on login node on cluster triggers predictions which almost kills login node)
3) Because the predictions are run during collection changes to the code might not retrigger the collection and thus new code will still use the old predictions, which creates a feeling of false safety
4) On some devices these full fledged tests fail to run, making tests not load at all
---------
Co-authored-by: Nathan Habib <nathan.habib@huggingface.co>
Co-authored-by: Nathan Habib <30601243+NathanHB@users.noreply.github.com>
