feat: expand Hugging Face model catalog

[alejandro-ao]

Description

Add 28 live-routable Hugging Face Inference Providers models from Pi's catalog, bringing Tau's built-in Hugging Face catalog to 46 models. Each model includes context limits, output limits, capabilities, compatibility flags, and pricing metadata. The existing default remains unchanged.

Three Pi entries that failed live routing were deliberately excluded: Qwen/Qwen3-Next-80B-A3B-Thinking, XiaomiMiMo/MiMo-V2-Flash, and moonshotai/Kimi-K2-Thinking.

Also add catalog coverage, user documentation, a development note, and a release note.

User experience

Hugging Face users can select the expanded, live-validated model set directly in /model, without maintaining a custom ~/.tau/catalog.toml overlay or seeing models that currently have no available inference provider.

Issue

No linked issue; addresses the reported Hugging Face model catalog gap.

Manual validation

1. Save a Hugging Face token with /login huggingface or set HF_TOKEN.
2. Open /model, filter to Hugging Face, and confirm the expanded model list appears.
3. Select a newly added model and send a short prompt.

Live validation used Tau's packaged catalog, saved Hugging Face credentials, and the prompt Reply with exactly: OK. All retained additions streamed content successfully except XiaomiMiMo/MiMo-V2.5-Pro, which reached the router but returned transient 429 Model busy responses.

Checks

• uv run pytest (1053 passed)
• uv run ruff check .
• uv run ruff format --check .
• uv run mypy
• uv build
• cd website && hugo --minify