better multi-gpu inference and training in the RL example

[rizar]

we will soon merge the RL examples #47 , and we will postpone the following work from there

• better use of multiple GPUs with vLLM for inference
• multi-GPU training using HF accelerate