Update models and address auto import of agents

[MarcCote]

This pull request introduces several improvements and refactorings related to language model (LLM) handling, agent registration, and token counting. The main focus is on making model handling more robust and maintainable by centralizing model lists, improving logic for model-specific parameters, enhancing token counting, and refining agent registration checks.

Model Handling Improvements:

• Centralized lists of supported models for Claude, OpenAI, and Gemini in both agents/llm.py and agents/reasoning.py, replacing scattered hard-coded lists with shared variables for easier maintenance and consistency. Logic throughout the agents now uses these lists to determine model-specific parameter handling. [1] [2] [3] [4] [5] [6]

Token Counting and Model Support:

• Expanded supported OpenAI models in tales/token.py to include new variants (gpt-5.1, gpt-5.2) and improved the token counting logic to handle boilerplate tokens more accurately. Also, Gemini token counter now strips the gemini/ prefix if present. [1] [2]

Agent Registration Robustness:

• Improved the register function in tales/agent.py to check for duplicate agent registration more accurately by comparing class names, preventing accidental overwrites or misleading errors. [1] [2]

Benchmarking and Logging:

• Added more informative logging when importing agent modules, aiding debugging and reproducibility. [1]