Refresh the models list, add some tooling for dynamically pulling the static list

[cmbrose]

Fixes #19 (as a side effect 😄)

(Generated by git diff forked/main parse_models_json.py refresh_models_json.py ../llm_github_models.py | llm -m github/o3-mini "Summarize this git diff into a PR body")

This PR makes several enhancements and additions to the GitHub models integration and associated tooling:

1. Changes in llm_github_models.py:
    • The chat models list (CHAT_MODELS) has been updated to include an extra field, api_version. Each model tuple now contains: model name, whether it supports streaming, input modalities, output modalities, and the API version (set to None by default or a preview version for select models like "o3", "o3-mini", and "o4-mini").
    • The register_models function now unpacks this extra api_version field and passes it to the GitHubModels constructor.
    • In the GitHubModels class, the api_version is stored and used to add an extra key in the client call, replacing the previous hardcoded logic for "o3-mini".
    • Minor formatting improvements (e.g., in the attachment_as_content_item and build_messages functions) help with clarity and maintainability.

2. Updates in tools/parse_models_json.py:
    • The script now categorizes models from models.json into “chat-completion”, “embeddings”, and “unknown” buckets.
    • For chat models, the tuple now includes the new api_version field determined via a helper function. For embeddings, an extra dimensions list is added (for example, [256, 1024] or [512] for specific models).
    • The function supports_streaming has been updated to mark additional models (such as "o3", "o4-mini") as not supporting streaming.
    • The script generates a Markdown fragment (“models.fragment.md”) that documents model details. It now produces separate sections for supported chat models and embedding models, including images, usage information, publisher details, and a formatted description.

3. New script: tools/refresh_models_json.py:
    • This new utility fetches the latest models catalog from the GitHub API using a valid GITHUB_TOKEN.
    • It builds URLs for model detail pages and scrapes each model’s JSON details (using BeautifulSoup) from the GitHub Marketplace.
    • The resulting list of models is sorted and written to a models.json file.
    • This script facilitates periodically refreshing local metadata without waiting for a live API.

Overall, these changes improve model metadata management by adding versioning support and richer attributes, automate documentation generation through Markdown fragments, and provide a new tool for updating the models catalog by combining API calls with web scraping.

[cmbrose]

Looking through the PRs, this definitely collides with #18 and #9. Closing this one out in favor of aligning with those others.