Create run_rl.py with ART RL loop

[saum7800]

A new file, dev/tau-bench/run_rl.py, was created, mirroring run.py but implementing an ART RL training loop.

Key changes and additions include:

• ART RL Training Loop: The core train function orchestrates the RL process, utilizing art.TrainableModel, art.gather_trajectory_groups, and model.train().
• Configuration: New Pydantic models, TauBenchTrainingConfig and TauBenchPolicyConfig, were introduced to manage RL-specific hyperparameters and integrate tau-bench's RunConfig.
• Trajectory Generation: The rollout_tau_bench_task function adapts tau-bench's task evaluation to generate art.Trajectory objects, converting agent interactions and rewards into a format suitable for ART.
• Agent Integration: The agent_factory from tau-bench.run is used, with the agent's internal model dynamically overridden to point to the art.TrainableModel's inference endpoint during rollouts.
• Argument Parsing: parse_args was extended to include RL-specific command-line arguments while retaining most run.py arguments for tau-bench compatibility.
• Evaluation: Periodic evaluation is performed using evaluate_model on a validation set, leveraging the same rollout_tau_bench_task mechanism.

The implementation reuses tau-bench abstractions like get_env and agent_factory to maintain consistency and minimize changes to the existing codebase.