Rename openrlhf to marti, add HF cache under repo, update deps (torch 2.9, vllm)

[deema-A]

Description

Updates the codebase to use the marti package name instead of openrlhf, improves the multi-agent training script for local/simple runs, and updates dependencies.

Changes

Package rename (openrlhf → marti)

• examples/multi-agent/run_train_chain.sh: Switched CLI and workflow paths from openrlhf to marti (e.g. openrlhf.cli.multi_agent_train_ppo_ray → marti.cli.multi_agent_train_ppo_ray, workflow/processor paths under marti/agent_workflows/).
• marti/trainer/ray/vllm_engine.py: Updated worker_extension_cls from openrlhf.trainer.ray.vllm_worker_wrap.WorkerWrap to marti.trainer.ray.vllm_worker_wrap.WorkerWrap.

Script / environment

• ROOT_DIR: If unset, derived from the script path so the script works from any working directory.
• Hugging Face cache: Set HF_DATASETS_CACHE and HF_HOME under the repo (e.g. ${ROOT_DIR}/.cache/huggingface) to avoid PermissionError when ~/.cache/huggingface isn’t writable (e.g. on shared systems).
• Defaults: MODEL_DIR="Qwen", MAX_LEN=30000 for the example run.

Dependencies (requirements.txt)

• Pinned torch to 2.9; added vllm >0.8.5.post1.
• Added: json5, latex2sympy2, pylatexenc, srsly, word2number.

Testing

• Multi-agent chain training script runs with the new paths and env.
• No remaining references to openrlhf in the updated code paths.