removed drop_last forwarding from make_eval_dataloader

[AlbinSou]

Before that, when calling

strategy.train(experience, eval_streams=[scenario.val_stream], drop_last=True)

The drop_last argument was used both in the training and in the evaluation dataloader. This is a quickfix to fix this behavior, it would be nice to find a better way of distinguishing kwargs that are forwarded to eval/train loaders, or even to functions outside of loaders.

[AntonioCarta]

looks like test_cumulative_accuracy is failing again

[AlbinSou]

looks like test_cumulative_accuracy is failing again

I know and I still have no clue why ... Still works perfectly locally (contrarily to the topk test which fails locally in my case). If only I could reproduce the error locally that would help a lot.

[AntonioCarta]

please try to fix this. You can try to create the same env that we use in the CI by creating a fresh environment (or using the docker image we use for tests).