improve client.zero_shot_classification()

[MoritzLaurer]

This PR does 3 things:

1. It adds support for the hypothesis_template argument from the 0-shot pipeline, which is important for customizing 0-shot classification
2. It removes the requirement for having at least 2 labels in the label list. This wasn't really necessary and there are cases where checking just one label with multi_label=True is useful.
3. It updates the doc strings and adds an example

Tested it with a HF endpoint running the model deberta-v3-large-zeroshot-v2

Example for use with hypothesis_template:

from huggingface_hub import InferenceClient

client = InferenceClient()

output = client.zero_shot_classification(
    model="https://h9qyt7jenlitt7j6.us-east-1.aws.endpoints.huggingface.cloud",
    text="I really like our dinner and I'm very happy. I don't like the weather though.",
    labels=["positive", "negative", "pessimistic", "optimistic"],
    multi_label=True,
    hypothesis_template="This text is {} towards the weather"
)

# output:
ZeroShotClassificationOutputElement(label='negative', score=0.9231801629066467),
ZeroShotClassificationOutputElement(label='pessimistic', score=0.8760990500450134),
ZeroShotClassificationOutputElement(label='optimistic', score=0.0008674879791215062),
ZeroShotClassificationOutputElement(label='positive', score=0.0005250611575320363)

Example for use with single label:

client = InferenceClient()

output = client.zero_shot_classification(
    model="https://h9qyt7jenlitt7j6.us-east-1.aws.endpoints.huggingface.cloud",
    text="I really like our dinner and I'm very happy. I don't like the weather though.",
    labels=["positive"],
    multi_label=True,
    hypothesis_template="This text is {} towards the weather"
)

# output
[ZeroShotClassificationOutputElement(label='positive', score=0.0001598795352037996)]

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Wauplin]

All 3 changes looks good to me. Thanks @MoritzLaurer for improving these!

I left a few minor comments. Could you also run make style locally and push the changes? Doing so is supposed to fix styling issues (e.g. line to long) + generate the exact same code for the _async client. Thanks in advance!

[MoritzLaurer]

Thanks for the review @Wauplin ! I've applied your suggestions and ran make style.
One weird thing is that running python utils/generate_async_inference_client.py --update removes references to async from the async client docstring (see this commit). I've first added references to async manually, but then the checks didn't pass anymore. Just flagging this in case that's unintended

[Wauplin]

Thanks for iterating @MoritzLaurer! Left a couple more comments. About the example issue, I will have a look at it. Don't worry about manually changing _async then. Just apply make style and I'll take care of it. Thanks!

[Wauplin]

Looks good to me! Thanks @MoritzLaurer for taking the time to also improve the docstrings :)

Some tests are failing but it's unrelated to your PR. I fixed the script that generates the _async code to properly transcribed the async examples (see f479f08).

[MoritzLaurer]

great, thanks a lot for your review @Wauplin !