feat: make llama.cpp a first-class local provider

[alejandro-ao]

Summary

• add a built-in llama-cpp provider with optional authentication and dynamic OpenAI model discovery metadata
• add tau llama-cpp setup to discover /v1/models and persist the server URL/model selection
• add tau llama-cpp doctor to probe connectivity, streaming, and tool-call support
• omit bearer authorization for OpenAI-compatible providers configured without authentication
• document the new local-model workflow and architecture boundary

Test plan

• uv run pytest
• uv run ruff check .
• uv run ruff format --check .
• uv run mypy
• cd website && hugo --minify
• cd website && npx --yes pagefind@latest --site public