refactor evaluator tests

[lvwerra]

Since IMDb is currently unavailable, also the evaluator tests have been failing (See #154). I took this as an opportunity to refactor the tests a bit:

• dataset: now the dataset is just a dummy dataset and does not rely on the hub.
• pipeline: in most places I replaced the pipeline with a dummy pipeline so we also don't rely on the hub. after all we want to test the evaluator and not loading the model from the hub. only where e.g. loading via string is tested a model is still downloaded but we could probably also mock this in the future.

Let me know what you think.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[ola13]

Looks great! Thanks @lvwerra