Fix InferenceEndpointsLLM not using cached token

[gabrielmbmb]

Description

InferenceEndpointsLLM was not automatically using the cached Hugging Face token (cached because huggingface-cli login).

In addition, this PR has updated the AsyncLLM.__del__ to just close the async loop if it was newly created by the AsyncLLM, otherwise it could raise a RuntimeError when calling close if the AsyncLLM didn't create it, because it's managed by other lib which might be still using it (such as pytest-asyncio).

[codspeed-hq]

CodSpeed Performance Report

Merging #690 will not alter performance

_{Comparing fix-inference-endpoints-llm-cached-token (27241da) with develop (918c19f)}

Summary

✅ 1 untouched benchmarks