Fix resolve chat completion URL

[Wauplin]

Fix #2484.

With the introduction of base_url in #2384 and a refacto in #2410, some bugs have been introduced on the logic to resolve the chat completion URL. This PR fixes definitely this with a dedicated _resolve_chat_completion_url method. I added a bunch of parametrized tests to check every possible cases. It is now possible to:

• pass a model id as InferenceClient(model=...)
• pass a model id as InferenceClient().chat_completion(model=...)
• pass a url as InferenceClient(model=...), InferenceClient(base_url=...) or InferenceClient().chat_completion(model=...)
  • the URL can have a trailing / or not
  • the URL can end with /v1 or not (for OpenAI compatibility)
  • the URL can end with /v1/chat/completions or not
  • the URL can be a local TGI instance or an Inference Endpoint url
• pass a url to InferenceClient and a model_id in chat_completion (already the case before)

With all of these use cases tested and working, I think this is fixed for good :)

---

Two minor unrelated changes:

1. some typing: ignore to fix CI (after Implemented https://github.com/huggingface/huggingface_hub/issues/2516 #2532)
2. Better error in src/huggingface_hub/utils/_http.py (after oversight in Refacto error parsing (HfHubHttpError) #2474)

Failing test is unrelated.

[Wauplin]

@MoritzLaurer FYI, I've made a full example for Inference Endpoints => InferenceClient and it works well:

from huggingface_hub import get_inference_endpoint


endpoint = get_inference_endpoint("smollm-360m-instruct-hgl")  # need to start an IE and copy its name
client = endpoint.resume().wait().client


output = client.chat_completion(
    [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Count to 10"},
    ]
)

print(output.choices[0].message.content)

[hanouticelina]

Looks good to me! thanks @Wauplin! Nice test coverage 😄

[Wauplin]

Thanks for the review! Failing tests are unrelated so I'll merge this one :)