Adding simple agent and option to call llm with tools

[MarcCote]

This pull request adds support for LiteLLM as a new language model backend in the codebase. It introduces a new LiteLLM class that integrates the LiteLLM library, enabling both tool-calling and non-tool-calling modes, and updates the model instantiation logic and imports accordingly.

LiteLLM integration:

• Added a new implementation LiteLLM in debug_gym/llms/litellm.py, supporting both tool-calling and non-tool-calling modes, with robust error handling, token counting, and response parsing logic.
• Updated the model instantiation logic in debug_gym/llms/base.py to allow selection of LiteLLM when "litellm" is present in the tags.
• Included LiteLLM in the __init__.py import and export lists for the llms module. [1] [2]

Agent improvements:

• Added a new SimpleAgent class in debug_gym/agents/simple_agent.py that parses tool calls from LLM responses using a custom XML-like format, executes them, and manages agent-environment interaction and logging.