feat: multi-agent environment integration

[Benjamin-eecs]

This pull request introduces a comprehensive integration of the TAU-BENCH Retail benchmark into the GEM (Gym for LLM Agents) framework, providing a clean, GEM-native implementation for evaluating LLM agents on realistic retail customer service tasks. The changes include new documentation, data loaders, tool interfaces, and evaluation scripts, as well as the addition of mock data assets and rule definitions to support the benchmark.

Key additions and improvements:

Documentation and Usage:

• Added a detailed README.md in both the root and tau_bench_retail directories, describing the integration, architecture, usage instructions, model support, and research experiment setup for the TAU-BENCH Retail benchmark in GEM. [1] [2]

Benchmark Assets and Data Handling:

• Introduced a new assets/data/__init__.py module for loading mock user, order, and product data from JSON files, supporting the benchmark's environment and tools.
• Added a readme.md in assets/data/ explaining the mock data generation process and schema design philosophy.
• Provided a .gitignore to exclude experiment results and Python cache files from version control.

Tooling and Evaluation Environment:

• Implemented a minimal Tool base class in assets/base_tool.py to standardize tool interfaces for the TAU-bench environment.
• Added assets/rules.py with explicit agent behavior and evaluation rules for the retail environment, ensuring agent actions are consistent with benchmark requirements.

These changes collectively enable robust, reproducible benchmarking of LLM agents in a realistic retail scenario, supporting multiple providers and facilitating research on user instruction clarity and tool-based reasoning.

[N00bcak]

Hi there, nice work!

There are just two conceptual details that I think may improve the design of the API a bit and I hope you can consider them here.

[lkevinzc]

LGTM, great work!