[InferenceClient] Add dynamic inference providers mapping

[hanouticelina]

Companion PR to huggingface/huggingface.js#1173. This is to give a first idea on how to implement the dynamic inference providers mapping before moving on with the TODOs.

TODOs: (a bit similar to the hf.js PR)

• handle prod/staging status -> warn the user when the model is in staging mode.
• handle if model supports both text-generation and conversational => not needed for now => let's make them "conversational"-only
• add and update tests.
• The HfApi().model_info call is cached.

[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.

[julien-c]

i'm wondering if, for Developer experience, we should still support an (empty by default) mapping that Partners can use locally before implementing the server-side mapping

[julien-c]

(and in the future let's add a link to some doc on how to implement the server-side mapping here)

[hanouticelina]

yes, I kept the hardcoded list

[Wauplin]

(updated the PR to align with hf.js client + did some refacto. Still some tests to fix I suppose ^^)

[Wauplin]

I've completed the PR. I still need to check a few things but it should be nearly done now. I took the opportunity to -once again 🙈- refactor how the providers are defined. We now use a base class that is inherited in all providers. Steps to prepare_request are split in atomic methods that can be tested individually (to prepare api_token, mapped_model, base_url, full url, payload, body, etc.). Goal is to reuse as much things as possible and therefore making the introduction of a new provider effortless. The case of LLMs is particularly interesting to optimize (see sambanova.py implementation).

I've also refactored the providers' tests to test only the parts that are overwritten in subclasses. The problem with current tests were that a lot was duplicated and it required the provider to be live on production (to get mapping). Since implementation is now centralized, we can also centralize the tests (still needs to be done). And testing atomic methods prevent from having to test end-to-end with production mapping.

[hanouticelina]

the refactored version is much much better, especially for testing ❤️ thanks!

[Wauplin]

Current state looks good to me :) @hanouticelina can you have a last look at it?

[hanouticelina]

All good on my side! some tests are failing but seem unrelated but i'm not sure if they are flaky or not, first time seeing connection errors from the model download cdn https://cdn-lfs-dev.hf.co

[Wauplin]

Let's merge for now :) I've reported internally (here) our CI issues but it's not related indeed.