Check for mismatch in device setup in evaluator

[mathemakitten]

Per #285, we missed an edge case in the compute function for the evaluator where if you pass a pipeline which was already instantiated without a device then it won't run on device even if a device is later passed to compute.

Alternatives considered were:

• Change nothing, but warn the user that they've passed a device to compute but their pipeline doesn't have a device attached. Bad because users can mindlessly scroll past it.
• Override the device used in the pipeline when a device has been detected but not used, and throw a different warning to let them know. Seems error-prone in multi-device setups, and reinstantiating already-instantiated pipes seems very bad.
  So I went with @lhoestq's suggestion to raise an error to force the user to disambiguate.

Tests with:

from datasets import Dataset, load_dataset
from evaluate import evaluator
from transformers import pipeline

data = load_dataset("Maysee/tiny-imagenet", split="valid", use_auth_token=True)

pipe = pipeline(
    task="image-classification",
    model="google/vit-base-patch16-224",
    # device=0  # If this commented it will trigger an error
    )

task_evaluator = evaluator("image-classification")
eval_results = task_evaluator.compute(
    model_or_pipeline=pipe,
    data=data,
    metric="accuracy",
    label_mapping=pipe.model.config.label2id,
    device=0
)

[HuggingFaceDocBuilderDev]

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

[lvwerra]

Looks great! Could you add a few tests to check if the desired behaviour is achieved?

[lhoestq]

Cool ! Raising an error is definitely ok in ambiguous situations

[lhoestq]

I think it's ok to raise an error is this case as well :)

[mathemakitten]

Done! Also added some tests :)

[lhoestq]

Thank you !