Fix chat completion token limit for GPT-5 models

[black-yt]

Summary

• choose the Chat Completions token-limit argument based on the model family
• use max_completion_tokens for GPT-5 and o-series models, including provider-prefixed IDs such as openai/gpt-5.1
• keep max_tokens for legacy chat-completions models
• add requests to package dependencies because the public structai import path imports structai.pdf
• bump the package version to 0.1.23 for the PyPI release
• add a no-network unittest suite covering public imports, LLM request kwargs, parsing helpers, IO round trips, Judge exact-match behavior, concurrency ordering, and utility functions
• add a GitHub Actions test workflow that runs the unittest suite on Python 3.10, 3.11, and 3.12

Testing

• PYTHONDONTWRITEBYTECODE=1 /home/xwh/miniconda3/envs/agent/bin/python -m unittest discover -s tests -p "test_*.py"