Add stop string support to rl.

[NicoGrande]

Description

Introduces support for stop string, which tell vLLM to stop generating as soon as any of the tokens included in this field are generated. This uses the new rollout_vllm_sampler_kwargs argument introduced here: google/tunix#1169

Tests

Command:

NEW_MODEL_DESIGN=1 TPU_BACKEND_TYPE=jax python3 -m src.maxtext.trainers.post_train.rl.train_rl src/maxtext/configs/post_train/rl.yml \
  model_name=gemma3-4b \
  tokenizer_path=google/gemma-3-4b-it \
  run_name=$WORKLOAD \
  base_output_directory=$OUTPUT_PATH \
  hf_access_token=$HF_TOKEN \
  batch_size=4 \
  num_batches=5 \
  scan_layers=False \
  hbm_utilization_vllm=0.4 \
  rollout_data_parallelism=2 \
  rollout_tensor_parallelism=2 \
  allow_split_physical_axes=true \
  load_parameters_path=$CHECKPOINT_PATH \
  vllm_hf_overrides='{architectures: ["MaxTextForCausalLM"]}' \
  vllm_additional_config='{"maxtext_config": {"model_name": "gemma3-4b", "log_config": "false"}}' 2>&1 | tee  grpo_out_gemma3.txt

Logs

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!