fix(plan): fetch exact Hugging Face repo IDs

[hannibal-lee]

What

• Fetch an exact Hugging Face repository when whichllm plan owner/repo cannot find it in the cached model catalog.
• Skip the bulk model catalog fetch for exact Repo ID queries.
• Reuse the existing model metadata parser and respect HF_ENDPOINT.
• Add clear errors for missing, private or gated repositories, network failures, and insufficient model metadata.
• Keep directly fetched repositories out of the shared ranking cache.

Why

whichllm plan documents support for Hugging Face Repo IDs, but it previously searched only the limited cached catalog. Valid repositories outside that catalog returned No model found.

This particularly affected ilsp/Llama-Krikri-8B-Instruct, which exists on Hugging Face but does not appear in the default model listing.

Fixes #163

Testing

• Tests pass (pytest) — 493 passed
• New tests added
• Tested on real hardware (not hardware-related)

Notes

Verified against the real Hugging Face API with:

whichllm plan ilsp/Llama-Krikri-8B-Instruct --json