llms/huggingface: update default inference endpoint to router.huggingface.co

[bozhouDev]

PR Checklist

• Read the Contributing documentation.
• Read the Code of conduct documentation.
• Name your Pull Request title clearly, concisely, and prefixed with the affected package (llms/huggingface: ...).
• Checked there isn't already a PR solving this the same way (searched open PRs; none touch the default URL).
• Provide a description / reference the issue it solves — Fixes #1428.
• Describes the source of new concepts.
• References existing implementations as appropriate.
• Contains test coverage for new functions — N/A (constant-only change; the existing TestHuggingFaceLLMStandardInference replay fixture is updated to cover the new endpoint).
• Passes all golangci-lint checks — go vet passes; I couldn't run golangci-lint locally, but this is a one-line constant value change with no new code.

Description

Fixes #1428.

HuggingFace deprecated https://api-inference.huggingface.co and requests to it now return 404 (the endpoint was retired on 2025-11-01). llms/huggingface still used it as defaultURL, so HuggingFace LLM and embedding calls fail by default unless the caller overrides the URL.

This points defaultURL at the hf-inference provider on the new router.

URL construction note: the client builds request URLs as fmt.Sprintf("%s/models/%s", c.url, model) (see internal/huggingfaceclient/inference.go and embeddings.go), so the base must omit a trailing /models. Using https://router.huggingface.co/hf-inference yields:

https://router.huggingface.co/hf-inference/models/<model>

which is HuggingFace's documented replacement for the legacy task-based endpoint. (Note: the literal https://router.huggingface.co/hf-inference/models suggested in the issue would produce a doubled path — .../hf-inference/models/models/<model> — so the base is set without the trailing /models.)

Verification

go build ./llms/huggingface/...   # ok
go vet   ./llms/huggingface/...    # clean
go test  ./llms/huggingface/...    # ok (both packages)

The TestHuggingFaceLLMStandardInference replay fixture is updated to target the new endpoint so the test stays consistent with the new default.

Caveat (please review): the replay fixture's recorded response is the prior 404 from the deprecated host (the test already t.Skips on 404). I don't have HuggingFace credentials to re-record a live 200 against the new endpoint, so a fresh recording with HF_TOKEN would be a welcome follow-up.